<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Robinson-27225</title>
    <description>The latest articles on DEV Community by Robinson-27225 (@robinson27225).</description>
    <link>https://dev.to/robinson27225</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4161771%2F29a5fec9-eb74-4d66-a7ba-9d2e105f630c.png</url>
      <title>DEV Community: Robinson-27225</title>
      <link>https://dev.to/robinson27225</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/robinson27225"/>
    <language>en</language>
    <item>
      <title>Tour of the Branches: a small open model that tells you who is buried around you, with no signal</title>
      <dc:creator>Robinson-27225</dc:creator>
      <pubDate>Fri, 30 Oct 2026 00:31:06 +0000</pubDate>
      <link>https://dev.to/robinson27225/tour-of-the-branches-a-small-open-model-that-tells-you-who-is-buried-around-you-with-no-signal-17kc</link>
      <guid>https://dev.to/robinson27225/tour-of-the-branches-a-small-open-model-that-tells-you-who-is-buried-around-you-with-no-signal-17kc</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week4-2026-10-26"&gt;Hacktoberfest Open-Source AI Challenge: Week 4&lt;/a&gt; ("Ghost in the Machine").&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;Old cemeteries have no signal and a lot of names. &lt;strong&gt;Tour of the Branches&lt;/strong&gt; is a web app that turns that into a walk through your family tree.&lt;/p&gt;

&lt;p&gt;Before you leave home you build a &lt;em&gt;tour pack&lt;/em&gt; from &lt;a href="https://www.wikitree.com" rel="noopener noreferrer"&gt;WikiTree&lt;/a&gt;, the free collaborative family tree: your ancestors, your relatives within a few degrees, or everyone buried in one cemetery. At the cemetery, GPS picks the nearest one, you tap a name, and a small open model, &lt;strong&gt;Gemma&lt;/strong&gt;, tells that person's story using only what their WikiTree profile says. It also says how they are related to you, links back to their profile, and can read the story aloud.&lt;/p&gt;

&lt;p&gt;Everything the story needs is already on the phone, so it works with no signal. Gemma writes the stories with Ollama on a laptop, or in the browser on the phone's own graphics chip. Nothing about your family or your location is sent anywhere to be processed.&lt;/p&gt;

&lt;p&gt;The ghosts in this machine are real people with sourced profiles. The machine is a model that has to be taught not to make things up about them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;"Genealogy without sources is mythology."&lt;/strong&gt; (1)&lt;/p&gt;

&lt;p&gt;Walking through a cemetery, or a town, and seeing how you are related to the people who are buried or lived there is powerful. Hearing their stories as you pass their headstones, or walk the paths they traveled, turns names into people and gives you a deeper understanding of your family history.&lt;/p&gt;

&lt;p&gt;That quote is also the rule this app follows. A story that drifts away from the record is mythology with a nicer voice, so every story here comes from a sourced WikiTree profile, and anything the record doesn't support is thrown away.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Try it:&lt;/strong&gt; &lt;a href="https://apps.wikitree.com/apps/robinson27225/branch-tour/index.html" rel="noopener noreferrer"&gt;Tour of the Branches&lt;/a&gt;. "Load the sample pack" works without a WikiTree login and with fictional people. Building a real pack needs a WikiTree login.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Video:&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/Ymun_k_yljU" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;The example I filmed is &lt;strong&gt;Molalla Memorial Cemetery&lt;/strong&gt; in Oregon. I typed its name and the app came back with:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;47 profiles found&lt;/strong&gt; in the category,&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;48 profiles&lt;/strong&gt; in the pack, &lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;45 with biographies&lt;/strong&gt;,&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;46 stories written&lt;/strong&gt; for the cemetery, every one checked against the record,&lt;/li&gt;
&lt;li&gt;every person named by how they are related to me, where that can be worked out.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Two stones made good examples. &lt;strong&gt;Herman Carl Harris (1922–2000)&lt;/strong&gt; comes out as &lt;em&gt;"your eighth cousin three times removed, through John Cooper (~1612–1689)"&lt;/em&gt;, and WikiTree's own Relationship Finder gives the same answer. &lt;strong&gt;Quinten Wyland (1919–2005)&lt;/strong&gt; was married twice. His first wife has a WikiTree profile; the second, Betty Lou, lies beside him and had none, so I made one. &lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/robinson27225" rel="noopener noreferrer"&gt;
        robinson27225
      &lt;/a&gt; / &lt;a href="https://github.com/robinson27225/branch-tour" rel="noopener noreferrer"&gt;
        branch-tour
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Tour of the Branches&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;Walk a cemetery with your family tree in your pocket. Before you leave home you build a &lt;em&gt;tour pack&lt;/em&gt; from &lt;a href="https://www.wikitree.com" rel="nofollow noopener noreferrer"&gt;WikiTree&lt;/a&gt;; at the cemetery GPS picks the nearest place, you tap a name, and a small open model, &lt;strong&gt;Gemma&lt;/strong&gt;, tells that person's story using only what their WikiTree profile says. It also says how they are related to you, links back to their profile, and can read the story aloud. Everything the tour needs is on the phone, so it works with no signal.&lt;/p&gt;
&lt;p&gt;Made for the DEV Hacktoberfest Open-Source AI Challenge, Week 4 ("Ghost in the Machine").&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Write-up:&lt;/strong&gt; &lt;a href="https://dev.to/robinson27225/tour-of-the-branches-a-small-open-model-that-tells-you-who-is-buried-around-you-with-no-signal-17kc" rel="nofollow"&gt;Tour of the Branches: a small open model that tells you who is buried around you, with no signal&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Video:&lt;/strong&gt; &lt;a href="https://www.youtube.com/watch?v=Ymun_k_yljU" rel="nofollow noopener noreferrer"&gt;Tour of the Branches at Molalla Memorial Cemetery&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Try it:&lt;/strong&gt; &lt;a href="https://apps.wikitree.com/apps/robinson27225/branch-tour/index.html" rel="nofollow noopener noreferrer"&gt;apps.wikitree.com/apps/robinson27225/branch-tour&lt;/a&gt; (the sample pack works without a WikiTree login)&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Plain JavaScript modules, no build step…&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/robinson27225/branch-tour" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;License: MIT.&lt;/p&gt;

&lt;p&gt;It is plain JavaScript modules with no build step. About 2.4 MB zipped, including the bundled WebLLM library. 146 tests.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;File&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/builder.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Builds a pack from the WikiTree API, WikiTree Plus and the category page&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/kinship.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Names the blood relationship between two people from their ancestor lines&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/narrate.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;The prompt, the plain-template fallback, the retries, and the checks on what the model wrote&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/quality.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Tells text from a loop of repeated words, for when a model breaks down&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/webllm.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Runs Gemma in the browser&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/ollama.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Setup help for running Gemma through Ollama instead&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/stories.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Saves stories to a file and loads them on another device&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/nearby.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Keeps only the people who rest within a distance of a place&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/offline.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;The "is this ready for no signal?" checklist&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/sort.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Orders the people at a stop&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/voice.js&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Optional ElevenLabs voice, with clips saved on the device&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. The pack
&lt;/h3&gt;

&lt;p&gt;WikiTree's API returns at most 1,025 profiles per page, and a page that hits that cap says "Maximum number of profiles (1025) reached." I first read that as a hard stop. It is not: the page still carries its profiles, and asking for the next page carries on. Following the pages, my seven-degree family circle comes back whole: 5, 12, 69, 333, 1,154, 2,281 and 4,126 profiles for one to seven degrees, and the full pack (4,032 people, 2.4 MB) builds in about half a minute. A WikiTree community member (Riel) pointed me at paging after I had wrongly told people the limit was a wall.&lt;/p&gt;

&lt;p&gt;A cemetery is a different problem: the API cannot list who is in a category. WikiTree Plus can. Its page address only fills in a form, though, and a script then calls a separate search when you press &lt;strong&gt;Get profiles&lt;/strong&gt;. I found the real request by reading that script. Now the app asks Plus who is in the cemetery, reads those profiles through the API, and checks each one's categories, because Plus matches words and not only categories.&lt;/p&gt;

&lt;p&gt;Plus also runs a day or more behind WikiTree. Herman Harris and Betty Lou Wyland were both in the cemetery category on WikiTree and missing from the app. So the build also reads the profile links on the category page, adds a spouse who is already in the category, and takes a typed list of WikiTree IDs for anyone else. None of those are trusted: every profile is checked for the category before it is kept.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Getting a small model to tell the truth
&lt;/h3&gt;

&lt;p&gt;This is where most of the work went. Gemma is fast and small, and it makes things up. Real examples from my testing:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Asked about my father's grave, it wrote &lt;em&gt;"a peaceful place to reflect … he's remembered as a father and a man who is now part of this community."&lt;/em&gt; The profile says none of that.&lt;/li&gt;
&lt;li&gt;It wrote &lt;em&gt;"He was a father to one child."&lt;/em&gt; I had given it the line &lt;em&gt;"Relationship to the visitor: your father"&lt;/em&gt; and it invented a count.&lt;/li&gt;
&lt;li&gt;For Benjamin Franklin Noyer it wrote a story that named a "Mary". Nothing in his record has a Mary.&lt;/li&gt;
&lt;li&gt;On my phone, a wrong window setting made it load without any error and then write &lt;em&gt;"John Cook John Cook John Cook…"&lt;/em&gt; down the screen, and a column of dashes. My fact checks passed that, because it contained no false facts. It was not language at all.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So the app does not trust the model. It checks twice. First, every story is checked against the record before you see it:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;// condensed from src/narrate.js
const FILLER = ['peaceful', 'remembered as', 'beloved', 'part of this community',
                'rest in peace', 'legacy', 'will be remembered'];
// a year, a name or place in mid-sentence, a count of children or siblings,
// or a stock phrase: if the record does not say it, the story is refused
export const isGrounded = (text, person, place) =&amp;gt; ungroundedWord(text, person, place) === '';
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Second, it checks that the text is a story at all: no word repeated four times in a row, no line repeated three times, no column of fragments.&lt;/p&gt;

&lt;p&gt;A refusal is not always right, and I had to fix the checks as often as the model. The check refused "November" because the word was not in the record, which only had "1933-11-13", so dates now go to the model as words. It refused "Let" at the start of a sentence. It refused &lt;em&gt;"one of six children"&lt;/em&gt; when the biography said &lt;em&gt;"one of 6 children"&lt;/em&gt;, until I made number words and digits the same number. Each fix made the check more exact and kept it strict about what the record does not say.&lt;/p&gt;

&lt;p&gt;When a story is refused, the model gets two more tries, each a little more careful. If all three fail, the app uses a plain narration built from the record alone, and the screen says why and shows what the model wrote. For Molalla's &lt;strong&gt;Margaret Luella Bagby&lt;/strong&gt; the plain version reads, in part: &lt;em&gt;"In 1860, 4 year old Margaret was one of 6 children living on the family farm at Molalla in Clackamas County, Oregon with parents William and Harriet."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Here is a story Gemma wrote through Ollama for &lt;strong&gt;Joanna Aileen Barstow&lt;/strong&gt;: &lt;em&gt;"Here, beside us, rests Joanna Aileen Barstow. She was born on June 25th, 1863, in Stafford, Oregon. Joanna married William Milton Bird in 1881…"&lt;/em&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Naming the relationship
&lt;/h3&gt;

&lt;p&gt;WikiTree's "N generations" turned out to be the steps up to a shared ancestor and back down. Me to Herman is 12 steps up on my side and 9 on his, which is 21, and which makes him an eighth cousin three times removed. I built &lt;code&gt;kinship.js&lt;/code&gt; on that and checked it against WikiTree's own list: eight relatives at Molalla could be traced, and all eight generation counts matched (14, 20, 20, 21, 22, 22, 23, 23). The rest are further out or related by marriage, and the app says "relationship not traced" for them and does not guess.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fovvnsf2qn9ctpu48h46k.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fovvnsf2qn9ctpu48h46k.png" alt="Relationship Finder" width="799" height="311"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Working with no signal, and where it broke
&lt;/h3&gt;

&lt;p&gt;The app is a service worker plus a few files. Stories and voice clips are saved the first time they are made, and the model weights are downloaded once (about 1 to 2 GB) and kept by the browser.&lt;/p&gt;

&lt;p&gt;Testing it in the field showed me what a demo hides. On Chrome for iPhone the app would not reload with no signal, because that browser cannot keep a web app's files for offline use. Safari with Add to Home Screen can. A weak signal is worse than none, because a request can hang for a minute instead of failing, so the service worker now gives up on the network after three seconds and uses its saved copy. A checklist on the Narrator tab ("Ready for no signal?") says what is and is not saved on the device.&lt;/p&gt;

&lt;p&gt;Then Gemma itself. On my iPhone, loading it made the page close and reload, twice in a row. I cannot see inside the phone, and the likely cause is memory. So the app now notes which step Gemma is on, and if the page closes during it, pauses Gemma, says which step it was on, and carries on with plain stories and the voice until I turn it back on.&lt;/p&gt;

&lt;p&gt;The fix that mattered more: a laptop running Gemma through Ollama writes better stories than the phone can. A button on each place writes every story, saves them to a file, and the phone loads the file. Each loaded story is checked against the record again before it is used. So the phone tells stories Gemma wrote, with no model running on the phone and no signal.&lt;/p&gt;

&lt;h3&gt;
  
  
  5. The voice
&lt;/h3&gt;

&lt;p&gt;The voice is optional and uses ElevenLabs with your own key. Only the finished story text leaves the device, and each clip is saved so it plays again with no signal. On an iPhone the browser only lets a page start sound during a tap, and a story takes seconds to write, so the app starts a silent sound inside the tap and plays the real clip through that same player later. Everything else stays local.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;It works where the signal doesn't.&lt;/strong&gt; A hosted model is useless in a cemetery in the hills. An open model that runs on your own laptop, or in a browser tab, is not.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It keeps family data private.&lt;/strong&gt; A pack holds the names, dates, places and relatives of a family, and the app knows where you are standing. With a local model none of that is sent to anyone to be processed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It can be moved.&lt;/strong&gt; The same Gemma runs in a phone's browser or through Ollama on a laptop. When the phone's browser could not hold the model, I moved the writing to the laptop and kept the stories. The pack and the checks stayed the same.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It can be checked.&lt;/strong&gt; Because I run the model myself, I could put my own checks around every sentence it writes, regenerate freely, and throw away what the record does not support. That is much harder to do when every retry has a price.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where open did not win.&lt;/strong&gt; The 1B model writes plainer, less fluent stories than I would expect from a large hosted model, I did not benchmark that, and on my phone it sometimes did not fit at all. Fluency was not the hard part, though. Faithfulness was, and that is where owning the model helped. The nicest-sounding part of the app, the voice, is a closed service, which is why it is optional.&lt;/p&gt;

&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;I built this with Claude Code over several days. Things worth saying honestly: the tests and a real browser caught mistakes the agent made (a field named &lt;code&gt;bio&lt;/code&gt; that it read as &lt;code&gt;Bio&lt;/code&gt;, so every biography was empty; an edit that deleted the Build button's handler while the syntax check still passed; a "write every story" button that stopped the whole run on one refusal). The reading of WikiTree Plus's script, the choice of what to trust, and what to refuse were mine.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Spirit of It
&lt;/h2&gt;

&lt;p&gt;Día de los Muertos keeps people present by saying their names. A cemetery does the same thing in stone, and a family tree does it in data. This app puts the two together: it says the names out loud, and it says only what is known.&lt;/p&gt;

&lt;p&gt;Betty Lou’s story is a good example. At first glance, the Betty beside Quinten looked like the Betty he married in 1940, already attached to his profile. A closer look showed that the first Betty’s birth and death dates don’t match the headstone, and Find a Grave has her buried in a different cemetery under another surname. So who is the Betty buried with Quinten in Molalla? A little research turned up a 1958 marriage record for Quinten (his second marriage) and a Betty Lou Orem. She now has a &lt;a href="https://www.wikitree.com/wiki/Orem-182" rel="noopener noreferrer"&gt;profile&lt;/a&gt; on WikiTree, and the photo of their shared headstone is linked to both of their profiles.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx06nkexgznle6q5pqctv.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx06nkexgznle6q5pqctv.png" alt="Quinten and Betty Lou Wyland's headstone, Molalla Memorial Cemetery" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Another example is Robert Churchill, who lived only a day and is remembered with a profile and ancestry on WikiTree&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frhogy942j0f5uaib5daa.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frhogy942j0f5uaib5daa.png" alt="Robert's profile on WikiTree: Churchill-1653" width="794" height="578"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Gemma:&lt;/strong&gt; the model that writes every story, in the browser and through Ollama.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;ElevenLabs:&lt;/strong&gt; the optional spoken voice, with clips saved for offline.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;(1) Schulze, Lorine McGinnis. "Genealogy without Sources is Mythology." &lt;em&gt;Olive Tree Genealogy&lt;/em&gt;. Accessed 3 Aug 2023.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>genealogy</category>
    </item>
    <item>
      <title>WikiTree PhotoTag: 2,500 faces named in 50 minutes, and not one photo leaves my computer</title>
      <dc:creator>Robinson-27225</dc:creator>
      <pubDate>Tue, 20 Oct 2026 20:31:13 +0000</pubDate>
      <link>https://dev.to/robinson27225/wikitree-phototag-2500-faces-named-in-50-minutes-and-not-one-photo-leaves-my-computer-3007</link>
      <guid>https://dev.to/robinson27225/wikitree-phototag-2500-faces-named-in-50-minutes-and-not-one-photo-leaves-my-computer-3007</guid>
      <description>&lt;p&gt;&lt;em&gt;This is my submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge: Week 3&lt;/a&gt;, theme "Slow and Steady".&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;My family album is 2,470 scanned photos in about two dozen folders: school portraits, 1940s snapshots, and plenty of "copy 2" duplicates. I also manage family profiles on WikiTree, and every one of those profiles would be better with a photo. But the album is full of people who are still alive, and putting their faces on a public website is not my call to make. So the photos have sat on my disk for years.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;WikiTree PhotoTag&lt;/strong&gt; does the dreaded part:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;It scans the album and finds every face.&lt;/li&gt;
&lt;li&gt;It groups look-alike faces, so I name a &lt;em&gt;group&lt;/em&gt; once instead of a face at a time.&lt;/li&gt;
&lt;li&gt;I tell it who each group is, using the people in my WikiTree &lt;strong&gt;CC7 export&lt;/strong&gt; (everyone within seven connections of me).&lt;/li&gt;
&lt;li&gt;It exports &lt;strong&gt;only the photos where every face is named and every person is deceased&lt;/strong&gt;, one folder per WikiTree profile, with a ready-to-paste caption like &lt;code&gt;[[Id|Name]] with [[Id|Name]], 1955.&lt;/code&gt;
&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Everything else stays on my machine. Uploading to WikiTree stays manual, which is how WikiTree wants it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;All screenshots are from my real album, taken with the tool's own &lt;code&gt;--blur&lt;/code&gt; mode. &lt;strong&gt;Every face is pixelated, and living relatives' names are hidden&lt;/strong&gt;, except for named deceased relatives of an older generation (the tool can keep anyone born after a year you choose hidden too).&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F88jkmay94pxzi42d37ke.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F88jkmay94pxzi42d37ke.jpg" alt="Naming screen with every face pixelated" width="800" height="605"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Name faces: each card is one group of look-alike faces. I click the right person, or one of the suggestions. The ✎ marks a scan with writing on it.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fnajnzwbh83bvsulttfn7.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fnajnzwbh83bvsulttfn7.jpg" alt="Named people tab" width="800" height="596"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Named people: everyone named so far, with deceased people marked as exportable. I can select single faces here and move them or send them back.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Foo4miete5fqmhp94ocq9.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Foo4miete5fqmhp94ocq9.jpg" alt="Find a person: unnamed faces ranked by how much they resemble a named relative" width="800" height="596"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Find a person: pick a relative and see every unnamed face the model thinks is them, best match first. (These candidates are pixelated because they are not named yet.)&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;A 19-person family photo. Four older relatives are named. Everyone else is pixelated, &lt;strong&gt;including faces the main detector missed&lt;/strong&gt;: blur mode runs a second, much more sensitive detector and hides anything face-like that is not a named deceased person.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fjw01kzqhrhvrqvgx0sn3.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fjw01kzqhrhvrqvgx0sn3.jpg" alt="Multi-person photo with four named relatives and every other face pixelated" width="800" height="553"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuzwqjb7t5acb4isxd4ox.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuzwqjb7t5acb4isxd4ox.jpg" alt="Export preview" width="720" height="270"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/robinson27225/wikitree-phototag" rel="noopener noreferrer"&gt;github.com/robinson27225/wikitree-phototag&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;MIT licensed, about 1,300 lines of Python, 25 tests. Setup is in the README: &lt;code&gt;init&lt;/code&gt;, &lt;code&gt;models&lt;/code&gt;, &lt;code&gt;scan&lt;/code&gt;, &lt;code&gt;cluster&lt;/code&gt;, &lt;code&gt;label&lt;/code&gt;, &lt;code&gt;export&lt;/code&gt;, plus optional &lt;code&gt;read-text&lt;/code&gt; and &lt;code&gt;xmp&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;The stack is deliberately small: &lt;strong&gt;OpenCV&lt;/strong&gt; for both models, &lt;strong&gt;SQLite&lt;/strong&gt; for the state, a &lt;strong&gt;stdlib HTTP server&lt;/strong&gt; with one HTML page for the UI. No GPU, no framework, no account.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Faces:&lt;/strong&gt; YuNet detects them and SFace turns each into a 128-number embedding. Both are small ONNX models from the OpenCV Zoo. The 2,470 photos produced 5,353 faces in well under ten minutes on a laptop. Nine byte-identical duplicate files were skipped automatically.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Grouping:&lt;/strong&gt; my first attempt was a greedy "assign to the nearest running average" pass. It turned 3,791 of 5,353 faces into one giant group, because a running average drifts toward a generic face and starts matching everyone. Switching to &lt;strong&gt;average-linkage hierarchical clustering&lt;/strong&gt; (SciPy) fixed it: the largest group is now 326 faces, and they are clearly the same person. Over-splitting is cheap here, because naming a person twice costs one extra click, while a merged pair of different people costs a lot of confusion. I picked the threshold with that in mind. A &lt;a href="https://dev.to/yakhyo/changing-a-face-recognition-model-also-changes-your-threshold-3p3n"&gt;dev.to post on how a face model change also changes your threshold&lt;/a&gt; (by yakhyo) was a useful reminder not to treat 0.36 as a magic number.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hints from the tree:&lt;/strong&gt; the CC7 gives names and dates. Album folders like &lt;code&gt;0A2 Carl William Robinson Sr&lt;/code&gt; and file names like &lt;code&gt;CarlSr_Fran_Teresa 003.jpg&lt;/code&gt; become suggestions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Narrowing by date:&lt;/strong&gt; only 8 of my 2,470 scans carry a year in their file name or EXIF, so I build a date &lt;em&gt;window&lt;/em&gt; for every photo from several sources. A year in the file name or EXIF is strongest. So is a &lt;strong&gt;printed date stamp&lt;/strong&gt; that Gemma reads off the print ("MAY 64" becomes 1964): that took me from 8 dated photos to &lt;strong&gt;58&lt;/strong&gt;. Next, the people already named in the photo: nobody is in a picture before they were born or after they died, so one named person bounds who else could be there. Last, a rough decade that Gemma guesses from clothing, hairstyles and photo quality. I tested that guess on the photos that do have a year: it was right for a 1907 and a 1969 photo, but 10–30 years off on several others. So it only adds a loose window (about 30 years either way) and never overrides the tree. Windows only &lt;em&gt;demote&lt;/em&gt; candidates ("dates don't fit"), they never hide anyone, because a wrong guess must not hide the right person.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Gemma reads the writing:&lt;/strong&gt; &lt;strong&gt;Gemma 3 (4B) running in Ollama&lt;/strong&gt; transcribes handwriting and printed captions, and the tool looks for relatives' full names in the result. I ran it over all 2,470 photos (about 2.5 seconds each, about two hours in total, entirely offline). &lt;strong&gt;257 had readable writing&lt;/strong&gt;, 184 of them photos with faces, for example &lt;em&gt;"for Grandpa Carl William Robinson / b. 11-13-1933"&lt;/em&gt;. A surprise: much of the text on scans without faces was &lt;strong&gt;documents&lt;/strong&gt; (certificates, discharge papers, an obituary), not photo backs, because my backs aren't scanned as separate files. The writing on the prints themselves was mostly names, signs, and lab date stamps. A ✎ badge in the UI marks scans that have writing on them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Find a person:&lt;/strong&gt; pick anyone from the CC7 and see every unnamed face the model thinks is them, best match first, then tick the right ones and accept. It is the quickest way to fill out a relative once a few of their faces are named. At a 50% match the top faces were clearly right, but around 40% many were someone else, so I accept by eye rather than by threshold.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Older groups first:&lt;/strong&gt; there is no face-age model, so an optional ordering uses black-and-white prints, early dates, the &lt;code&gt;Ancestors/&lt;/code&gt; folder and deceased folder names to put probably-deceased groups at the top. That is where the exportable photos are.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Set aside:&lt;/strong&gt; classmates and strangers get one click ("not family"), and for a school photo, one click sets aside every face still unnamed in it. They are never matched again, and any photo containing one stays private, since a stranger could be living.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Non-destructive, and XMP tags:&lt;/strong&gt; originals are never touched. Exports are copies with an &lt;code&gt;.xmp&lt;/code&gt; sidecar holding standard face regions, the people in the image, a caption and the WikiTree IDs. A separate &lt;code&gt;phototag xmp --write&lt;/code&gt; command can tag the album itself with a &lt;code&gt;&amp;lt;photo&amp;gt;.jpg.xmp&lt;/code&gt; beside each original (1,503 of my photos have named faces), so the names show up in digiKam and similar tools. It is a dry run by default and never overwrites a sidecar that already exists.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  The part that matters: the privacy gate
&lt;/h3&gt;

&lt;p&gt;The CC7 has no "living" flag, so the tool treats someone as deceased only if they have a death date, or were born before 1910. Then the export rule is a few lines:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;unknown&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;sum&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;f&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;real&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;person_id&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt;
&lt;span class="n"&gt;others&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;sum&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;f&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;real&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;person_id&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;OTHER&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;living&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;pid&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;pid&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;people&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="n"&gt;roster&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;is_deceased&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;pid&lt;/span&gt;&lt;span class="p"&gt;)]&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;unknown&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;reason&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;unnamed&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;unknown&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt; face(s) not yet named&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
&lt;span class="k"&gt;elif&lt;/span&gt; &lt;span class="n"&gt;others&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;reason&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;other&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;others&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt; face(s) set aside as not family (could be living)&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
&lt;span class="k"&gt;elif&lt;/span&gt; &lt;span class="n"&gt;living&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;reason&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;living&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;possibly living: &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;, &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;living&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;elif&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="n"&gt;include_auto&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="nf"&gt;any&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;label_src&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;auto&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;f&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;real&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;reason&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;auto&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;contains names nobody confirmed (auto-accepted)&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
&lt;span class="k"&gt;else&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;export&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;A photo leaves only if no face is unnamed, no face is a stranger, nobody is possibly living, and every name was confirmed by a human. Unknown classmates in a school photo keep the whole photo private.&lt;/p&gt;
&lt;h3&gt;
  
  
  What went wrong (and changed the design)
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Folder names are hints, not facts.&lt;/strong&gt; A folder named for a relative contained mostly photos of children. A group of them was auto-suggested as that relative. Names guessed by the tool are now marked &lt;em&gt;auto&lt;/em&gt; and &lt;strong&gt;never exported&lt;/strong&gt; unless I pass &lt;code&gt;--include-auto&lt;/code&gt;. A wrong guess could otherwise file a living child under a deceased person's profile.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;First names lie.&lt;/strong&gt; Two family members share a name, and the living one is not in my CC7. Weak first-name matches now get a dashed "(weak)" button, and picking a deceased person asks for confirmation. I made exactly this mistake once while labeling (I named a living relative as the deceased one), so the &lt;strong&gt;Named people&lt;/strong&gt; tab now lets me select individual faces and move them, or send them back to unnamed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The detector misses faces.&lt;/strong&gt; The privacy gate only knows about faces that were detected. The README says so, and the demo blur uses the more sensitive second pass for that reason.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A tidy-up broke the app.&lt;/strong&gt; I cleared my Downloads folder and the CC7 file the tool pointed at was gone. Now &lt;code&gt;init&lt;/code&gt; keeps its own copy, and a missing file gives a clear message instead of a crash. My labels were never at risk, because they live in a separate database.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;My own screenshots nearly leaked.&lt;/strong&gt; The first blur mode hid faces but still printed living relatives' names on screen, and it showed a relative who died this year. Blur mode now hides living names, and &lt;code&gt;--clear-born-before&lt;/code&gt; keeps younger deceased relatives pixelated too.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Because the data is the problem. The photos show real people, many of them alive, and my relatives did not sign up to be uploaded to a cloud vision API. With open models, the whole pipeline (detection, recognition, caption reading, even the era guess) runs on a laptop with the network off.&lt;/p&gt;

&lt;p&gt;Open also meant I could change the parts that mattered. I could swap the clustering method when the first attempt collapsed, set my own match threshold, run a second detector purely to over-blur, treat Gemma's era guess as a loose hint after measuring how wrong it can be, and read the CC7 file as plain data. A hosted face-tagging service decides all of those for you, and a "this is a family member" feature would have been the opposite of what I needed: a way to be &lt;em&gt;conservative&lt;/em&gt;, not convenient.&lt;/p&gt;

&lt;p&gt;There's a community example of the same instinct: &lt;a href="https://dev.to/swmcc/indexatron-teaching-local-llms-to-see-family-photos-3npa"&gt;Indexatron&lt;/a&gt; by swmcc, which uses local models to analyse family photos for the same reason.&lt;/p&gt;
&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;I built WikiTree PhotoTag with Claude Code. This is a curated copy of that session: my prompts and its replies, with tool output left out and the names of living and recently deceased relatives removed.&lt;/p&gt;


&lt;div class="ltag-agent-session"&gt;
  &lt;div class="agent-session-header"&gt;
    
    &lt;span class="agent-session-tool-icon-badge" title="Claude Code"&gt;
&lt;/span&gt;
    &lt;span class="agent-session-title"&gt;Building WikiTree PhotoTag: a local-first family photo tagger&lt;/span&gt;
  &lt;/div&gt;

  &lt;div class="agent-session-scroll"&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;I have a few ideas and want to see if they are a fit for the week 3 challenge for hacktoberfest&lt;br&gt;
&lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;https://dev.to/challenges/hacktoberfest-week3-2026-10-19&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;One is a way to tag photos in a family album on your computer with their wikitree id. &lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;cc7 extract from wikitree&lt;/li&gt;
&lt;li&gt;photo album folder&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;since some of the photos may be of living people these would need to be kept on the computer. then the program could run and output any photos of deceased relatives to a folder ready for import to their wikitree profiles.&lt;/p&gt;

&lt;p&gt;capable of:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;tagging multiple people in a photo&lt;/li&gt;
&lt;/ul&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;&lt;strong&gt;Yes, it fits Week 3 ("Slow and Steady"), as long as the AI does the tagging rather than you.&lt;/strong&gt; The theme asks for open-weight or open-source AI that does a task you dread. Tagging a shoebox of scanned family photos is that kind of task. The challenge is &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge: Week 3&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What the challenge needs&lt;/strong&gt;&lt;br&gt;
- It has to be a new project built during the window (Oct 19–25).&lt;br&gt;
- Open-source or open-weight AI has to be what makes it work.&lt;br&gt;
- Judging puts writing quality first, then relevance to the theme, creativity, technical execution, and optional partner tech.&lt;br&gt;
- The post needs the tags &lt;code&gt;devchallenge&lt;/code&gt; and &lt;code&gt;hf26challenge&lt;/code&gt;.&lt;br&gt;
- The post must explain why open innovation matters, and has an optional "Lazy Payoff" section for the time saved.&lt;br&gt;
- The deadline is unclear. The page says Oct 25, 11:59 PM PDT, but the API's end time is Oct 25 06:59 UTC, which is Oct 24 11:59 PM PDT. Treat Oct 24 as the real deadline so you don't miss it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why your idea works&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;It has to run locally.&lt;/strong&gt; Photos of living relatives can't go to a cloud API. That gives you a strong "why open matters" answer, because the data never leaves the machine.&lt;br&gt;
- &lt;strong&gt;The lazy payoff is easy to measure.&lt;/strong&gt; You can say something like "label 40 faces once and 2,000 photos get tagged."&lt;br&gt;
- &lt;strong&gt;Multi-person tagging fits face clustering.&lt;/strong&gt; The program detects every face in a photo, groups similar faces across the album, and you name each group once. Open models like InsightFace or a local vision model can do this.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;@"~/Downloads/CC7_Robinson-27225_2026-10-20_15-40-02_all.csv"&lt;br&gt;
Risk 3 - yes it is a manual process.&lt;br&gt;
Risk 4 - see attached cc7 csv file for my profile&lt;/p&gt;

&lt;p&gt;Yes to all ideas to make it stronger&lt;/p&gt;

&lt;p&gt;Let's proceed with this one&lt;/p&gt;

&lt;p&gt;~/Pictures/Family&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Folder's set up (&lt;code&gt;~/code/Photo Identifier Tool&lt;/code&gt;, empty). I'll set up a Python environment and check that the face-detection stack installs, then build the tool.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Still building: the core modules are in (CC7 roster, scanner, clustering, export gate). Next I'm writing the labeling page and the command-line entry point.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Gemma is reading the faceless scans in the background (about 3 s each). While it runs, I'm adding filename and caption hints to the cluster suggestions.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Gemma has read 180 of 299 scans so far (29 had text). I'm drafting the README while it finishes.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The core pipeline works end to end on your real album. I'm pausing until the Gemma caption-reading pass finishes, and I'll look at its results then.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What's built&lt;/strong&gt; in &lt;code&gt;~/code/Photo Identifier Tool&lt;/code&gt;:&lt;br&gt;
- &lt;strong&gt;Scan:&lt;/strong&gt; I scanned your album, 2,470 photos. It found 5,353 faces and skipped 9 duplicate files.&lt;br&gt;
- &lt;strong&gt;Cluster:&lt;/strong&gt; &lt;code&gt;phototag cluster&lt;/code&gt; grouped the faces into 1,023 groups. The biggest groups were clearly single people when I looked at a sample. The groups of young children and some older women were mixed, which is what the "split off" button handles.&lt;br&gt;
- &lt;strong&gt;Hints:&lt;/strong&gt; The folder hint correctly flagged your own folder. File names like &lt;code&gt;CarlSr_Fran_Teresa 003.jpg&lt;/code&gt; also feed hints.&lt;br&gt;
- &lt;strong&gt;Label page:&lt;/strong&gt; The page runs at &lt;code&gt;http://&lt;span class="agent-session-redacted"&gt;[REDACTED]&lt;/span&gt;:8765&lt;/code&gt; and works. I tested labeling and export on a copy of the data, so none of your real labels were touched.&lt;br&gt;
- &lt;strong&gt;Export gate:&lt;/strong&gt; A photo is exported only if every face is named and every named person is deceased. 11 tests cover it.&lt;br&gt;
- &lt;strong&gt;CC7 data:&lt;/strong&gt; 4,069 of the 4,131 people count as deceased. The other 62 are withheld as possibly living. Living siblings in your folders (several siblings) mostly aren't in the CC7 at all, so those faces stay unnamed and their photos stay private, which is the right behavior.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Decisions for you&lt;/strong&gt;&lt;br&gt;
1. &lt;strong&gt;Demo screenshots.&lt;/strong&gt; The post can't show your family album, since many of the faces are living people. I'd suggest a demo with public-domain historical photos, or one that blurs faces. Which do you prefer?&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;a demo of mine with blurred photos for any that may be living, if possible. &lt;br&gt;
is there a better name for the tool? do you have any suggestions. something with wikitree in it since that is a part of it. wikitree phototag?&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The blank page is fixed. The labeling page now loads again at &lt;code&gt;http://&lt;span class="agent-session-redacted"&gt;[REDACTED]&lt;/span&gt;:8765&lt;/code&gt;; please reload it. The cause was my old server process running stale code after I changed the code. I restarted it with the current code and checked the page.&lt;/p&gt;

&lt;p&gt;Your two other questions:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Name.&lt;/strong&gt; I'd go with &lt;strong&gt;WikiTree PhotoTag&lt;/strong&gt;, as you suggested. It says what it does and the command stays &lt;code&gt;phototag&lt;/code&gt;. Other options I considered:&lt;br&gt;
- &lt;strong&gt;Shoebox to WikiTree&lt;/strong&gt;: friendly, but it doesn't say photos or tagging.&lt;br&gt;
- &lt;strong&gt;WikiTree Faces&lt;/strong&gt;: short, but sounds like it posts faces somewhere.&lt;br&gt;
- &lt;strong&gt;PhotoRoots for WikiTree&lt;/strong&gt;: catchy, but "PhotoRoots" may already be taken by other projects.&lt;/p&gt;

&lt;p&gt;If you use WikiTree in the name, add a line to the README saying it's an unofficial tool, not affiliated with WikiTree. Tell me whether to rename the project and README.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Blurred demo.&lt;/strong&gt; I've built it:&lt;br&gt;
- &lt;code&gt;phototag label --blur&lt;/code&gt; pixelates every face except people you've named as deceased. That includes the thumbnails and the full-photo view.&lt;br&gt;
- There's a new "Named people" tab, and the photo view draws names under the faces that are shown.&lt;br&gt;
- &lt;code&gt;phototag accept --hints 0.7&lt;/code&gt; can pre-name groups where the album folder, file name or caption agrees.&lt;/p&gt;

&lt;p&gt;To make the screenshots, I'd run the blur mode on a copy of your data and name a few deceased-relative groups there, such as the Peasley and Harrison ancestors. Your real labels stay yours to do. Should I go ahead with that?&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;yes, rename WikiTree PhotoTag&lt;br&gt;
yes, go ahead with the demo&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The project is renamed to &lt;strong&gt;WikiTree PhotoTag&lt;/strong&gt; (command still &lt;code&gt;phototag&lt;/code&gt;, with an "unofficial, not affiliated with WikiTree" note in the README). The blurred demo is built.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Demo images&lt;/strong&gt; are in &lt;code&gt;docs/&lt;/code&gt; and embedded in the README:&lt;br&gt;
1. The naming screen with every face pixelated.&lt;br&gt;
2. The "Named people" tab showing four deceased relatives clearly.&lt;br&gt;
3. The export preview: 14 photos of 4 deceased relatives would be exported, and 2,157 stay private.&lt;br&gt;
4. A 19-person family photo with one deceased man named and everyone else pixelated.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Check before posting:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Faces I verified by eye.&lt;/strong&gt; The demo names came from full-name file names such as &lt;code&gt;Alice Hester Harrison 001.jpg&lt;/code&gt;, and I looked over the faces. Two exported files look worth a second look: &lt;code&gt;Alice_Paul.jpg&lt;/code&gt; and &lt;code&gt;img172 copy.jpg&lt;/code&gt;, both filed under Alice Hester Harrison.&lt;br&gt;
- &lt;strong&gt;Deceased faces are clear in screenshots 2 and 4.&lt;/strong&gt; That is what you asked for, but confirm you're comfortable posting them.&lt;br&gt;
- &lt;strong&gt;Blur covers detected faces only.&lt;/strong&gt; My first blurred version missed two faces in the group photo, so I added a second, more sensitive pass that blurs anything face-like that isn't a named deceased person. The README now says to skim exported photos before uploading.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Two safety fixes I made along the way:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Folder hints are weak.&lt;/strong&gt; One group was filed under a man born in 1935, but its faces were 1970s–80s children. So names guessed by &lt;code&gt;seed&lt;/code&gt; or &lt;code&gt;accept&lt;/code&gt; are now marked auto and are left out of &lt;code&gt;export&lt;/code&gt; unless you pass &lt;code&gt;--include-auto&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;made a mistake, this is identified as the wrong naomi robinson. should be [living relative], living&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;I'll fix that now. I've backed up the database first. Then I'll make mistakes like this easy to undo and harder to make.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Fixed. Those 166 faces are now named &lt;strong&gt;[living relative]&lt;/strong&gt;, as a living, local-only person. A local-only name is never exported, so those photos stay private. Export preview now shows 0 photos ready, and 122 withheld as living. I backed up your database first (&lt;code&gt;data/phototag.backup-*.db&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why it happened.&lt;/strong&gt; The suggestion button offered "Naomi Lucy Robinson" because file names starting with &lt;code&gt;Naomi_…&lt;/code&gt; matched the one Naomi in the CC7. Your sister isn't in the CC7, since she's living, so the tool picked the wrong one. That's the same weak first-name hint problem I found earlier, and I should have flagged it in the UI instead of only in &lt;code&gt;accept&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What I changed so it's harder to repeat and easy to undo:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Weak hints are marked.&lt;/strong&gt; A first-name-only suggestion now has a dashed orange outline, says "(weak)", and has a tooltip explaining that the person may be a living relative who isn't in the CC7.&lt;br&gt;
- &lt;strong&gt;Deceased names get a warning.&lt;/strong&gt; Any suggestion or search result for a deceased person shows "deceased". Clicking it asks you to confirm that these photos could become exportable.&lt;br&gt;
- &lt;strong&gt;Undo on the "Named people" tab.&lt;/strong&gt; Each person now has two buttons:&lt;br&gt;
  - "Wrong person — send back to unnamed"&lt;br&gt;
  - "Move to someone else…" (WikiTree id or a plain name for a living relative)&lt;/p&gt;

&lt;p&gt;Please reload the page at &lt;code&gt;http://&lt;span class="agent-session-redacted"&gt;[REDACTED]&lt;/span&gt;:8765&lt;/code&gt; to pick up the new buttons.&lt;/p&gt;

&lt;p&gt;The Naomi cluster itself was a clear single-person group, so the clustering was fine. It was the name that was wrong.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;there are two carl william robinson's... my father and my brother. not sure which button is for which one. for deceased can the year range be added?&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Every suggestion button now shows the life years and WikiTree id, for example &lt;code&gt;Carl William Robinson (1933–2003) Robinson-27265&lt;/code&gt;. Living people on the list show "(living / local only)". Hard-reload the page (Cmd+Shift+R) to see it. Your screenshot is from the older page, which is also why the first-name hints in it don't show the "(weak)" tag.&lt;/p&gt;

&lt;p&gt;I checked the CC7, and there is only one Carl William Robinson in it: your father, 1933–2003, Robinson-27265. Your brother isn't in the CC7, because he's living. So the two Carl buttons in your screenshot should be:&lt;br&gt;
- &lt;strong&gt;"Carl William Robinson · first name in file name"&lt;/strong&gt; is a weak first-name match to your father.&lt;br&gt;
- &lt;strong&gt;"Carl William Robinson · 65%" and "· 39%"&lt;/strong&gt; are face-similarity suggestions from people you've already named. If one of them was your brother, you'd have named him locally, and he'd now show as "(living / local only)".&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;This group is probably your brother, not your father.&lt;/strong&gt; The faces look like a man around 1969, in photos from different decades, and the file name hint only matches a first name. If it's your brother, use "Living / not on WikiTree" and type his name. That keeps those photos private. If it's your father, pick the &lt;code&gt;(1933–2003) Robinson-27265&lt;/code&gt; button, and you'll get a confirmation prompt because he's deceased.&lt;/p&gt;

&lt;p&gt;For two relatives with the same name, you'll now be able to tell them apart by life years. Two living people with the same name, like your brother and a nephew, you'd tell apart by typing distinct local names, for example "Carl William Robinson Jr". I haven't added a "Jr" suffix to the local-name entry itself.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;2853 of 5353 still to go. seems to work really well&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Those are strong numbers for the post. Here is where you stand.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Faces named&lt;/td&gt;
&lt;td&gt;2,500 of 5,353 (47%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Naming decisions you made&lt;/td&gt;
&lt;td&gt;81 (one per group)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;People named&lt;/td&gt;
&lt;td&gt;58 (10 deceased relatives, the rest living or local-only)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Photos ready to export&lt;/td&gt;
&lt;td&gt;70, covering those 10 deceased relatives&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Photos withheld as living&lt;/td&gt;
&lt;td&gt;737&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Photos withheld with unnamed faces&lt;/td&gt;
&lt;td&gt;1,362&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Roughly 31 faces per click works out to the "lazy payoff". The 2,853 faces left are the long tail of small groups, which will take much less naming for far fewer faces.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The export is small so far.&lt;/strong&gt; 70 photos is low because photos with any unnamed or living face stay private. As you name more, mostly the ancestors, it will grow. Naming the small groups in &lt;code&gt;Ancestors/&lt;/code&gt; will probably add the most exportable photos, since those are the people WikiTree profiles want.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;To finish by the deadline&lt;/strong&gt; (Oct 24 to be safe):&lt;br&gt;
1. &lt;strong&gt;Keep naming&lt;/strong&gt; for as long as you like. Anything left over simply stays private.&lt;br&gt;
2. &lt;strong&gt;How long did the 81 decisions take you?&lt;/strong&gt; I'll put the real time in the post. A rough number is fine.&lt;br&gt;
3. &lt;strong&gt;Run &lt;code&gt;phototag export&lt;/code&gt;&lt;/strong&gt; and skim the exported photos, since the gate can't see faces the detector missed.&lt;br&gt;
4. &lt;strong&gt;Commit and push to a public GitHub repo.&lt;/strong&gt; The challenge needs a code link, and the &lt;code&gt;data/&lt;/code&gt;, &lt;code&gt;out/&lt;/code&gt; and model files are already git-ignored. Tell me the repo name and whether to use your GitHub account, and I'll commit and push.&lt;br&gt;
5. &lt;strong&gt;Draft the DEV post&lt;/strong&gt; from the submission template (What I Built, Demo, Code, How I Built It, Why Open Innovation Matters, Lazy Payoff), saved as a private draft for you to review and publish.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;we need a way to indicate faces to not try to match again. for instance, classmates in a school photo&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Good idea. I'll add a "set aside" status. Those faces drop out of grouping and suggestions, and any photo containing one stays private, because an unknown classmate could be living. I'll also add a one-click "set aside everyone else in this photo" for school-photo situations.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Added: reload the page at &lt;code&gt;http://&lt;span class="agent-session-redacted"&gt;[REDACTED]&lt;/span&gt;:8765&lt;/code&gt; to see the new buttons on each group.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Not family: whole group.&lt;/strong&gt; Use this when a whole group is people you'll never name, such as a class of strangers.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Not family: selected.&lt;/strong&gt; Click the faces that aren't family, then use this button. That works when a group is mixed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Set aside everyone else in these photos.&lt;/strong&gt; This is the school-photo case. Name your relative's face first. Then in any other group, click one classmate's face and use this button. Every face still unnamed in that photo is set aside in one go. It asks you to confirm first.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;How set-aside faces behave:&lt;/strong&gt;&lt;br&gt;
- They leave the unnamed pile and are never grouped or matched again. They are also never used as examples for suggestions.&lt;br&gt;
- &lt;strong&gt;Any photo with one of them is withheld from export.&lt;/strong&gt; An unknown classmate could be living, so the tool keeps the photo private. The export preview shows how many were held back for this.&lt;br&gt;
- You can undo it on the "Named people" tab. There's a "Not family / unknown (set aside)" card at the bottom with a "send back to unnamed" button. The header now also shows a "set aside" count.&lt;/p&gt;

&lt;p&gt;I tested it on a copy of your data, so your real labels are untouched. Two new tests cover the rule that set-aside faces keep a photo private and are never regrouped.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;can you tell when i sent the first issue/question about the two naomi's?&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Your first Naomi message was sent at &lt;strong&gt;9:21 AM PDT on Oct 20&lt;/strong&gt; (16:21 UTC). I read this from the session transcript's timestamps.&lt;/p&gt;

&lt;p&gt;Your other recent messages were:&lt;br&gt;
- the two Carl William Robinsons question, at 9:42 AM PDT&lt;br&gt;
- "2853 of 5353 still to go", at 10:11 AM PDT&lt;/p&gt;

&lt;p&gt;If you want the time for the post's "lazy payoff" section, the tool doesn't record when each label was made, so I can't work out how long the naming took. I can add a timestamp to every label so the final numbers are exact, for example "N faces named in M minutes". It only takes effect from now on, so earlier labels would still be estimates. Want me to add it?&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;is there a way to have an option to show the older/probably decease faces first?&lt;br&gt;
yes, so the 50 minutes was the length of time spent&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Two things: (1) an "older / probably deceased first" ordering, and (2) timestamps on labels, so the 50 minutes can be recorded properly from here on. The tool has no face-age estimator, so I'll build the ordering from clues it does have: black-and-white look, known years, &lt;code&gt;Ancestors/&lt;/code&gt; folders and deceased hints.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Both are in, and 14 tests pass. Reload the page at &lt;code&gt;http://&lt;span class="agent-session-redacted"&gt;[REDACTED]&lt;/span&gt;:8765&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Older-first ordering.&lt;/strong&gt; The "Name faces" tab now has a dropdown in its intro line: &lt;em&gt;biggest first&lt;/em&gt; or &lt;em&gt;older / probably deceased first&lt;/em&gt;. The page remembers your choice. The tool has no face-age model, so groups are ranked from clues it does have:&lt;br&gt;
- black-and-white prints&lt;br&gt;
- an early year in the photo's file name or date&lt;br&gt;
- anything under &lt;code&gt;Ancestors/&lt;/code&gt;&lt;br&gt;
- an album folder named for someone deceased&lt;/p&gt;

&lt;p&gt;On your album, the 1907 &lt;code&gt;RobinsonReunion&lt;/code&gt; photos now come first. Faded colour scans of grandparents may rank lower than they should. If the order looks wrong in places, tell me what you see and I'll adjust the weights.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Timestamps.&lt;/strong&gt; Every label from now on is timestamped. The header will show "N min naming", counting only active time (gaps over 5 minutes aren't counted).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Your 50 minutes.&lt;/strong&gt; Your earlier 2,500 faces have no timestamps, so I'll record that part as your own figure for the post: &lt;strong&gt;81 naming decisions, about 2,500 faces, about 50 minutes&lt;/strong&gt;. The new timer only covers work from here on.&lt;/p&gt;

&lt;p&gt;Do you want me to commit and push the code to GitHub now, since the post needs a code link?&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;also need a way to move one face to someone else in a group on the named tab&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;also need a way to move one (or more) faces to someone else in a group on the named tab&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;You can now move individual faces on the &lt;strong&gt;Named people&lt;/strong&gt; tab. Reload the page to see it. I clicked through the page to check it. Selecting a face brought up the "1 selected" bar, and nothing in your data changed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How it works:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Select faces.&lt;/strong&gt; Click any faces to select them. Each person's card starts with 14 faces, and "Show all N faces" loads the rest so you can check the whole set.&lt;br&gt;
- &lt;strong&gt;Move them.&lt;/strong&gt; A bar appears with the count and these options:&lt;br&gt;
  - &lt;strong&gt;Move selected to…&lt;/strong&gt; Search the CC7 by name or id and click the right person. Results show the life years and WikiTree id.&lt;br&gt;
  - &lt;strong&gt;Use typed name (living / not on WikiTree).&lt;/strong&gt; This names them locally, so those photos stay private.&lt;br&gt;
  - &lt;strong&gt;Not family.&lt;/strong&gt; This sets the faces aside, like the classmates case.&lt;br&gt;
  - &lt;strong&gt;Send selected back to unnamed.&lt;/strong&gt; The faces return to the unnamed pile.&lt;br&gt;
- &lt;strong&gt;Deceased targets ask for confirmation.&lt;/strong&gt; The prompt says that photos where everyone is named and deceased become exportable.&lt;br&gt;
- &lt;strong&gt;Open the photo.&lt;/strong&gt; Each face has a ↗ link that opens its photo with everyone boxed. Use it to check a face before you move it.&lt;br&gt;
- &lt;strong&gt;Whole-person undo.&lt;/strong&gt; The old "send all back" button is still there, now labeled "Whole person is wrong".&lt;/p&gt;

&lt;p&gt;The app now has 16 tests. Two new ones cover moving some faces between people, sending faces back, and refusing an unknown WikiTree id.&lt;/p&gt;

&lt;p&gt;Do you want me to commit and push to GitHub now? I'd still need the repo name.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;commit and push to github. i believe the name should be: wikitree-phototag right?&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Yes, &lt;code&gt;wikitree-phototag&lt;/code&gt; is a good name. The code is committed and pushed, and the repo is &lt;strong&gt;private&lt;/strong&gt; for now: &lt;a href="https://github.com/robinson27225/wikitree-phototag" rel="noopener noreferrer"&gt;robinson27225/wikitree-phototag&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Pushed:&lt;/strong&gt; 26 files, the source, tests, README, MIT license and the four blurred demo images. I checked what was staged before committing. Not included: your photos, the database, thumbnails, exports, your CC7 file, the face models (anyone can fetch them with &lt;code&gt;phototag models&lt;/code&gt;) and the virtual environment.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;I made it private because making it public is your call.&lt;/strong&gt; The challenge needs a public link, so before you flip it, please look at &lt;code&gt;docs/demo-2-named-people.jpg&lt;/code&gt;. It shows the unblurred faces of several deceased relatives, and it's the one image I'd double-check. Also confirm the MIT license line, which names you as the copyright holder (I used your git name).&lt;/p&gt;

&lt;p&gt;When you're ready, this makes it public:&lt;/p&gt;

&lt;pre class="highlight bash"&gt;&lt;code&gt;gh repo edit robinson27225/wikitree-phototag &lt;span class="nt"&gt;--visibility&lt;/span&gt; public &lt;span class="nt"&gt;--accept-visibility-change-consequences&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;p&gt;I can run it for you if you say so.&lt;/p&gt;

&lt;p&gt;The challenge closes Oct 24 to be safe. Tell me when to start the DEV post draft. I have these ready for it:&lt;br&gt;
- 2,500 faces named in 81 decisions, about 50 minutes&lt;br&gt;
- 58 people named&lt;br&gt;
- 70 photos exportable so far, 737 withheld as living&lt;br&gt;
- the Gemma caption reading&lt;br&gt;
- the two mislabeling mistakes, which are a good honest story about why guesses never auto-export&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;demo image is good. all deceased.&lt;br&gt;
start the post draft&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;DEV rejected the GitHub embed because the repo is still private. I'll swap it for a plain link, and I'm also removing one line I invented ("a hundred copies of the same Christmas picture").&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The draft is saved to your DEV drafts: &lt;strong&gt;&lt;a href="https://dev.to/robinson27225/wikitree-phototag-2500-faces-named-in-50-minutes-and-not-one-photo-leaves-my-computer-3fo6-temp-slug-1027854/edit"&gt;edit it here&lt;/a&gt;&lt;/strong&gt;. It's marked as AI-assisted (&lt;code&gt;some_ai&lt;/code&gt;), the honest level for a post I drafted. Nothing is published. You publish from DEV, or tell me to.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Title:&lt;/strong&gt; WikiTree PhotoTag: 2,500 faces named in 50 minutes, and not one photo leaves my computer. Tags: &lt;code&gt;devchallenge&lt;/code&gt;, &lt;code&gt;hf26challenge&lt;/code&gt;, &lt;code&gt;python&lt;/code&gt;, &lt;code&gt;privacy&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;It follows the challenge's submission template: What I Built, Demo, Code, How I Built It, Why Open Innovation Matters, The Lazy Payoff and Prize Categories. The writing is first person, as you. I kept the three mistakes in because they show why guesses never auto-export: the clustering that collapsed into one giant group, the folder hint that mislabeled children, and the first-name mix-up. I linked community posts on face-model thresholds and local family-photo AI.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Before you publish, please check these:&lt;/strong&gt;&lt;br&gt;
1. &lt;strong&gt;Make the repo public.&lt;/strong&gt; The post links to it, and the demo images load from its raw GitHub URLs, so they show as broken until it's public. Say the word and I'll run the command to flip it.&lt;br&gt;
2. &lt;strong&gt;Refresh the numbers.&lt;/strong&gt; They're from just now: 75 exportable photos for 15 deceased relatives, and 737 withheld as living. They will change as you keep naming.&lt;br&gt;
3. &lt;strong&gt;Facts I took from you or inferred.&lt;/strong&gt; Please correct anything wrong:&lt;br&gt;
   - "I manage family profiles on WikiTree"&lt;br&gt;
   - the story that you named a living relative as the deceased one&lt;br&gt;
   - the 50 minutes and 81 decisions&lt;br&gt;
   - "well under ten minutes" for the scan&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;repo is now public&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The repo is public and checks out:&lt;br&gt;
- &lt;strong&gt;Repo:&lt;/strong&gt; &lt;a href="https://github.com/robinson27225/wikitree-phototag" rel="noopener noreferrer"&gt;robinson27225/wikitree-phototag&lt;/a&gt;, MIT license detected, and your local code is in sync with it.&lt;br&gt;
- &lt;strong&gt;Demo images:&lt;/strong&gt; all four load from GitHub (HTTP 200), so they'll render in the post.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;To finish:&lt;/strong&gt;&lt;br&gt;
1. &lt;strong&gt;Open the draft and read it through.&lt;/strong&gt; It still holds the numbers from earlier today and the facts I asked you to confirm.&lt;br&gt;
2. &lt;strong&gt;Preview the images.&lt;/strong&gt; Check the post preview on DEV to make sure all four screenshots show up.&lt;br&gt;
3. &lt;strong&gt;Optional polish:&lt;/strong&gt; now that the repo is public, I can swap the plain repo link in the Code section for DEV's GitHub embed card (the one DEV rejected earlier). Say so and I'll update the draft.&lt;br&gt;
4. &lt;strong&gt;Tell me when you're ready,&lt;/strong&gt; and I can publish it for you, or you can hit publish on DEV. Either way, aim for before the Oct 24 deadline.&lt;/p&gt;

&lt;p&gt;If you keep naming faces first, tell me, and I'll update the "75 photos / 15 relatives" numbers in the post right before you publish.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;what about a way to focus in on a certain person. for instance, it's prompting that it thinks the face is someone, a way to select from the cc7 that person and see all faces that it is proposing are of that person&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Good idea. I'm adding a &lt;strong&gt;"Find a person"&lt;/strong&gt; tab. Pick someone from your CC7, or someone you've named locally, and it lists every unnamed face the model thinks is that person, best matches first. You tick the right ones and accept them in one go.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;"Find a person" is added, and I tried it on your data. Reload &lt;code&gt;http://&lt;span class="agent-session-redacted"&gt;[REDACTED]&lt;/span&gt;:8765&lt;/code&gt;; it's the new tab next to "Named people".&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How to use it&lt;/strong&gt;&lt;br&gt;
1. &lt;strong&gt;Search for the person.&lt;/strong&gt; Type a name. It searches your CC7 and also the living relatives you've named only locally, and the results show life years and WikiTree id.&lt;br&gt;
2. &lt;strong&gt;See the candidates.&lt;/strong&gt; You get a grid of every still-unnamed face the model thinks is that person, best match first. Each face shows its match percentage, and ↗ opens the full photo.&lt;br&gt;
3. &lt;strong&gt;Pick the right ones.&lt;/strong&gt; Click faces to tick them, or use "Select all shown above N%". The "Minimum match" slider widens or narrows the list.&lt;br&gt;
4. &lt;strong&gt;Accept.&lt;/strong&gt; "Accept selected as…" names exactly the faces you ticked. If the person is deceased, it asks you to confirm first.&lt;/p&gt;

&lt;p&gt;By default it hides faces from photos dated outside the person's lifetime. You can untick that option.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What I saw:&lt;/strong&gt; I tried it on your father. He has 77 faces named already, and 34 unnamed faces came up at 40% or higher. The top 10 or so (the bearded man in glasses, around 50%) look right. By the 40–45% range, many are clearly someone else, including a young woman, so don't use "select all" at the low end. Raising the threshold to about 55–60% is safer. As a fix, I'd tick faces you recognize rather than accept everything above a number.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Limit:&lt;/strong&gt; it compares against faces already named as that person, so someone with no named faces shows nothing yet. In that case it tells you to name a group or two on the "Name faces" tab first.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;we need to do the ideas to make it stronger&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;narrow candidates by date&lt;/li&gt;
&lt;li&gt;read the writing on the back&lt;/li&gt;
&lt;li&gt;write xmp people tags&lt;/li&gt;
&lt;/ul&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;You're right that those three are only partly done. Here's what's missing:&lt;br&gt;
- &lt;strong&gt;Date narrowing:&lt;/strong&gt; it only works for the few photos that have a year in the file name or EXIF.&lt;br&gt;
- &lt;strong&gt;Reading the writing:&lt;/strong&gt; it only ran on scans without faces, and the text is only used as a hint.&lt;br&gt;
- &lt;strong&gt;XMP tags:&lt;/strong&gt; they're only written for exported copies, not for the album itself.&lt;/p&gt;

&lt;p&gt;I'll look at the data first so I fix the real gaps.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The three improvements are built and tested, but &lt;strong&gt;the app is down until your CC7 file is back&lt;/strong&gt;. I need your help with that.&lt;/p&gt;
&lt;h2 id="agent-session-1077-42-action-needed-the-cc7-file-is-missing"&gt;Action needed: the CC7 file is missing&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;CC7_Robinson-27225_…_all.csv&lt;/code&gt; is no longer in &lt;code&gt;~/Downloads&lt;/code&gt;. It looks like the folder was tidied around 10:41, and I can't find it anywhere else (Spotlight and Trash included). I restarted the labeling server, and it couldn't load the file, so the page at &lt;code&gt;http://&lt;span class="agent-session-redacted"&gt;[REDACTED]&lt;/span&gt;:8765&lt;/code&gt; is down. &lt;strong&gt;Your labels are safe in the database.&lt;/strong&gt; Only the people list is missing.&lt;/p&gt;

&lt;p&gt;Please do one of these:&lt;br&gt;
- Put the file back, or tell me where it went.&lt;br&gt;
- Re-export it from WikiTree's CC7 tool. Dates and names may have changed slightly, but your WikiTree IDs are the same, so it will work with your existing labels.&lt;/p&gt;

&lt;p&gt;Then run this (or tell me the path and I will):&lt;/p&gt;

&lt;pre class="highlight bash"&gt;&lt;code&gt;.venv/bin/phototag init &lt;span class="nt"&gt;--cc7&lt;/span&gt; /path/to/CC7.csv &lt;span class="nt"&gt;--photos&lt;/span&gt; ~/Pictures/Family
&lt;/code&gt;&lt;/pre&gt;

&lt;p&gt;I fixed the root cause: &lt;code&gt;init&lt;/code&gt; now keeps its own copy in &lt;code&gt;data/cc7.csv&lt;/code&gt;, which is git-ignored, so tidying Downloads can't break the app again. If the file is missing, the app now says what to do instead of crashing.&lt;/p&gt;
&lt;h2 id="agent-session-1077-42-the-three-improvements"&gt;The three improvements&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;1. Narrowing candidates by date.&lt;/strong&gt; Only 8 of your 2,470 photos have a year, so the old version rarely helped. A photo's possible date window now comes from three sources, strongest first:&lt;br&gt;
- &lt;strong&gt;A year in the file name or EXIF.&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Who is already named in the photo.&lt;/strong&gt; Someone born in 1933 can't be in a photo from 1920, and nobody appears after they died.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;file is back&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The CC7 file is restored and everything is running again. &lt;code&gt;init&lt;/code&gt; reported the same 4,131 people, so your labels line up. It also kept a local copy in &lt;code&gt;data/cc7.csv&lt;/code&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The page is back&lt;/strong&gt; at &lt;code&gt;http://&lt;span class="agent-session-redacted"&gt;[REDACTED]&lt;/span&gt;:8765&lt;/code&gt;. Your 2,497 named faces, 63 people and 22 set-aside faces are all intact.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The background pass has started again.&lt;/strong&gt; Gemma is reading captions and guessing an era for the ~2,170 photos with faces. It should take about 90 minutes. It saves in batches of 20, and the first batch hasn't landed yet, so I can't confirm numbers yet. It can run alongside your labeling.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;XMP preview:&lt;/strong&gt; a dry run says 1,503 photos have named faces, and none have sidecars yet. Nothing was written. Run &lt;code&gt;phototag xmp --write&lt;/code&gt; when you want those &lt;code&gt;.xmp&lt;/code&gt; files created beside your originals, or tell me to.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Once the Gemma pass finishes, the date narrowing will start using the rough eras too. Until then it uses file-name years and who's already named in each photo. I'll check back on it. If you'd like, I can commit and push the new work, then update the DEV post's "How I Built It" with the date narrowing and XMP changes.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;commit, push, and update the draft post&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;The draft has one edit of yours that I'll keep (you dropped the "unofficial" parenthetical in the first paragraph). I'm updating the rest from that version.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;I've pushed the code to GitHub and updated the DEV draft. Nothing is published.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Pushed:&lt;/strong&gt; commit &lt;code&gt;0349bdf&lt;/code&gt; is on &lt;code&gt;main&lt;/code&gt; at &lt;a href="https://github.com/robinson27225/wikitree-phototag" rel="noopener noreferrer"&gt;robinson27225/wikitree-phototag&lt;/a&gt;. The README is updated for all the new features. The 22 tests pass. I checked what was staged first, and none of your photos, database or CC7 file went up.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Draft, &lt;a href="https://dev.to/robinson27225/wikitree-phototag-2500-faces-named-in-50-minutes-and-not-one-photo-leaves-my-computer-3fo6-temp-slug-1027854/edit"&gt;edit here&lt;/a&gt;:&lt;/strong&gt; your edit to the first paragraph, dropping the "unofficial" parenthetical, is kept. Changes I made:&lt;br&gt;
- &lt;strong&gt;Narrowing by date:&lt;/strong&gt; a new section covering the three date sources. It includes the honest result of my test of Gemma's era guesses (right on two photos, 10–30 years off on several others), which is why they're only a loose window.&lt;br&gt;
- &lt;strong&gt;Gemma and the writing:&lt;/strong&gt; now notes that most readable text was documents, not photo backs, and that the pass was extended to photos with faces.&lt;br&gt;
- &lt;strong&gt;New bullets:&lt;/strong&gt; Find a person, Set aside, and XMP tags (the &lt;code&gt;phototag xmp&lt;/code&gt; command, 1,503 photos have named faces).&lt;br&gt;
- &lt;strong&gt;New "what went wrong" item:&lt;/strong&gt; the Downloads tidy-up that deleted the CC7 file.&lt;br&gt;
- &lt;strong&gt;Code and prize sections:&lt;/strong&gt; 1,300 lines, 22 tests, and the Gemma prize wording now mentions the era guess.&lt;/p&gt;

&lt;p&gt;The post is marked AI-assisted (&lt;code&gt;some_ai&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Two things to check before you publish&lt;/strong&gt;&lt;br&gt;
1. &lt;strong&gt;The Gemma pass on photos with faces is still running&lt;/strong&gt; (about 90 minutes total). The post says it ran, so wait until it finishes, or tell me and I'll adjust the wording. I'll check its progress whenever you ask.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;ol&gt;
&lt;li&gt;check back on it. did it take this long before?&lt;/li&gt;
&lt;li&gt;that's fine. it was by accident that it got selected with other files&lt;/li&gt;
&lt;/ol&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;&lt;strong&gt;1. The pass had crashed, and I've restarted it.&lt;/strong&gt; It stopped at 580 of 2,171 photos with "database is locked". The cause was my own design: it and your labeling page both write to the same database, and the page won a collision. Your labels were not affected. Here is what changed:&lt;br&gt;
- &lt;strong&gt;Fix:&lt;/strong&gt; the database now waits up to 60 seconds for a lock and uses a mode built for two writers. I've also restarted the labeling page, which is up (it shows your 2,497 named faces).&lt;br&gt;
- &lt;strong&gt;Resume:&lt;/strong&gt; the pass restarted and picks up where it stopped, with 1,591 photos left. At about 2.5 seconds each, that is roughly an hour.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Did it take this long before?&lt;/strong&gt; The speed per photo is the same as the first run. That earlier run took roughly 12–15 minutes for 299 scans with no faces. This one covers 2,171 photos, about seven times as many, so about 90 minutes in total. So far, 580 photos have an era guess, mostly the 1970s–2000s, and 56 photos with faces have writing on them.&lt;/p&gt;

&lt;p&gt;I'll check on it again whenever you ask. I'd suggest not publishing until it finishes, so the post's claim about it is true.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Good.&lt;/strong&gt; The Downloads line stays in the post as written.&lt;/p&gt;

&lt;p&gt;I haven't pushed the database fix yet. I'll include it in the next commit.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;we need logo and banner images, what do you suggest?&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Good first draft. Two flaws: the "deceased ✓" label on the banner is partly covered by a silhouette, and the green check on the logo covers the "g" in PhotoTag. I'm fixing both.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;I've made first drafts so you can react to something real. They're in &lt;code&gt;docs/branding/&lt;/code&gt;, and I've sent you the three images.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The idea:&lt;/strong&gt; the logo and banner show a family group photo where every head is pixelated except one named, deceased person. That person has green face brackets and a "deceased ✓" label. It's the tool's core idea (the privacy gate) in a single picture. The green check badge on the logo says "approved to leave", and a cream and sepia palette makes it feel like an old album. It uses no real faces or photos, and it doesn't copy WikiTree's own logo or branding, which fits an unofficial tool.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What's there:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Banner, 1000×420:&lt;/strong&gt; the DEV post cover image. The text reads "WikiTree PhotoTag. Name every face once. Only the deceased leave your computer."&lt;br&gt;
- &lt;strong&gt;Logo, 512×512:&lt;/strong&gt; a Polaroid with one figure inside face brackets, a second one pixelated, and the check badge.&lt;br&gt;
- &lt;strong&gt;GitHub social preview, 1280×640:&lt;/strong&gt; the same banner at the size GitHub uses for link previews.&lt;/p&gt;

&lt;p&gt;Everything is regenerated by &lt;code&gt;docs/branding/make_branding.py&lt;/code&gt;, so colors, text and sizes are easy to change.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How you could use them:&lt;/strong&gt; upload the banner as the post's cover image on DEV. Add the logo to the top of the README. Set the social preview in the GitHub repo's Settings, which only you can do in GitHub's web page.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Options if you want a different direction:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Typography only:&lt;/strong&gt; just the name in the serif font with a green face-bracket mark. Simpler, but less distinctive.&lt;br&gt;
- &lt;strong&gt;A real demo photo as the banner:&lt;/strong&gt; the blurred 19-person photo. It's authentic, but busier, and it shows your family.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;use wikitree font calibri and use brand colours&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;You already have the brand colors and Calibri installed, so I'll switch the drafts to those. I'll keep the "unofficial" label and I won't use WikiTree's logo.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;I've restyled all three images with your WikiTree brand colors and Calibri, and sent them to you.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Colors used, from your palette:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Dark green&lt;/strong&gt; &lt;code&gt;#25422D&lt;/code&gt; for the title, text and the named person.&lt;br&gt;
- &lt;strong&gt;Orange&lt;/strong&gt; &lt;code&gt;#f37c26&lt;/code&gt; for "PhotoTag" and the face brackets.&lt;br&gt;
- &lt;strong&gt;Yellow&lt;/strong&gt; &lt;code&gt;#FCB815&lt;/code&gt; for the "deceased ✓" label.&lt;br&gt;
- &lt;strong&gt;Lime and yellow greens&lt;/strong&gt; (&lt;code&gt;#8FC641&lt;/code&gt;, &lt;code&gt;#A5D167&lt;/code&gt;, &lt;code&gt;#E1F0B4&lt;/code&gt;, &lt;code&gt;#FFE270&lt;/code&gt;, &lt;code&gt;#FAD158&lt;/code&gt;) for the pixelated heads and the logo badge. The pixelated faces are now a green and yellow mosaic.&lt;br&gt;
- &lt;strong&gt;Pale green&lt;/strong&gt; &lt;code&gt;#E5EECF&lt;/code&gt; and &lt;code&gt;#E1F0B4&lt;/code&gt; for the background.&lt;/p&gt;

&lt;p&gt;I didn't use black from the palette. The text is Calibri, which is installed on your Mac. If you want the exact &lt;code&gt;#000000&lt;/code&gt; somewhere, tell me where.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What I left out:&lt;/strong&gt; I didn't use WikiTree's actual logo or tree mark, and the banner still says "UNOFFICIAL". With WikiTree's colors and font, it will look closer to official, so the "unofficial" label matters more now. You know WikiTree's brand rules better than I do, so you're the best judge of whether this is fine for a community tool. If it's not, I can switch to a neutral palette quickly.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Using them:&lt;/strong&gt; upload &lt;code&gt;banner-dev-1000x420.png&lt;/code&gt; as the cover image of the DEV post. &lt;code&gt;social-preview-1280x640.png&lt;/code&gt; goes in the GitHub repo's Settings, under "Social preview". That upload is a web-only step I can't do for you. &lt;code&gt;logo.png&lt;/code&gt; goes at the top of the README, and I can add it.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;nice looks good&lt;br&gt;
remove unoffical&lt;br&gt;
we'll be linking this and sharing on wikitree as a community tool&lt;/p&gt;

&lt;p&gt;commit and push&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The unofficial wording is gone, the logo is in the README, and everything is pushed. The latest commit is &lt;code&gt;96fb6a2&lt;/code&gt; on &lt;code&gt;main&lt;/code&gt;. All 22 tests pass.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Banner:&lt;/strong&gt; the strip now reads "OPEN MODELS · 100% LOCAL · COMMUNITY TOOL".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Everywhere else:&lt;/strong&gt; I took "unofficial" out of the README (which now says "A community tool for WikiTree family historians"), the package description and the command-line help. That also removed the "not affiliated with WikiTree" disclaimer. If you want any disclaimer back, say so.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pushed too:&lt;/strong&gt; the images (PNG and SVG) and &lt;code&gt;make_branding.py&lt;/code&gt; for regenerating them. The database fix that lets your labeling and the background pass run at once is in this commit as well.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Your DEV draft&lt;/strong&gt; already had that wording removed by you, so it needs no change.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Gemma pass:&lt;/strong&gt; it's running again and not crashing. It has 800 photos done and 1,371 to go, about an hour. The draft says it covered photos with faces, so I'd wait until it finishes before publishing.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Before you share it on WikiTree:&lt;/strong&gt;&lt;br&gt;
1. &lt;strong&gt;Set the cover image:&lt;/strong&gt; upload &lt;code&gt;banner-dev-1000x420.png&lt;/code&gt; as the draft's cover on DEV, or tell me and I'll set it from its GitHub link.&lt;br&gt;
2. &lt;strong&gt;Set the social preview:&lt;/strong&gt; in the repo's Settings, under "Social preview", upload &lt;code&gt;social-preview-1280x640.png&lt;/code&gt;. That has to be done on GitHub's website.&lt;br&gt;
3. &lt;strong&gt;Refresh the numbers in the post&lt;/strong&gt; (75 photos, 15 relatives) after your last naming session. Tell me when and I'll update them.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;need a new set of folders due to the buttons across the top changing&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;I'm reading "a new set of folders" as a new set of &lt;strong&gt;demo screenshots&lt;/strong&gt;, since the old ones show the old row of buttons (they're missing "Find a person"). One problem first: your real data now contains living people's names, and the old blur mode only hid faces, not names. I'll make blur mode hide living people's names too, then regenerate the images.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;sorry, mean new set of screenshots&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Got it, new screenshots. I'll go ahead, including the change so blur mode also hides living people's names. Your real data now has some, and a screenshot must not show them.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Pushed (commit &lt;code&gt;31f7d0d&lt;/code&gt;). Now I'm updating the draft with the new images and numbers. I'll fetch its current text first, in case you've edited it.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The new screenshots are done and in the repo and the draft, with the code pushed (commit &lt;code&gt;31f7d0d&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;New set in &lt;code&gt;docs/&lt;/code&gt;:&lt;/strong&gt;&lt;br&gt;
1. &lt;strong&gt;Name faces:&lt;/strong&gt; every face pixelated, with suggestions reading "Living relative (hidden)".&lt;br&gt;
2. &lt;strong&gt;Named people:&lt;/strong&gt; your older generation of deceased relatives, with the new face-select buttons and the full row of tabs.&lt;br&gt;
3. &lt;strong&gt;Find a person:&lt;/strong&gt; Alice Hester Harrison's candidates, ranked with match percentages.&lt;br&gt;
4. &lt;strong&gt;The 19-person photo:&lt;/strong&gt; now with four named relatives, rather than one.&lt;br&gt;
5. &lt;strong&gt;Export preview:&lt;/strong&gt; 81 photos of 15 deceased relatives, with 739 withheld as living.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Two things I changed so the screenshots are safe:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Names are hidden too.&lt;/strong&gt; My first attempt would have shown living relatives' &lt;em&gt;names&lt;/em&gt; on screen, because your data now has some. Blur mode now hides those names and the living people's cards.&lt;br&gt;
- &lt;strong&gt;Only the older generation is shown clearly.&lt;/strong&gt; The first Named people screenshot included [recently deceased relative] and others who died recently, which seemed wrong for a public post. A new option, &lt;code&gt;--clear-born-before 1950&lt;/code&gt;, keeps everyone younger pixelated. If you'd prefer a different cutoff, or to show them, say so. Your father (b. 1933) is still shown clearly, and I didn't ask before choosing that, so please check you're comfortable with it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A real bug I found while doing this:&lt;/strong&gt; two thumbnails showed as broken because dozens of parallel requests shared one database connection. I added a lock, so it's fixed. Nobody would have noticed it in normal use, but it could have affected the labeling page.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;yes, my father is fine as example&lt;br&gt;
i'll update dev by hand&lt;br&gt;
update github as well. &lt;br&gt;
waiting to publish until it's done&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;GitHub is updated.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Code and images:&lt;/strong&gt; they were already pushed (&lt;code&gt;31f7d0d&lt;/code&gt;), and the repo is in sync with your machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Repo description:&lt;/strong&gt; it still said "unofficial". It now reads: &lt;em&gt;"WikiTree PhotoTag: a local-first community tool that names the faces in your family album once, then exports photos of deceased relatives for their WikiTree profiles. Open models, nothing uploaded."&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Topics added:&lt;/strong&gt; &lt;code&gt;wikitree&lt;/code&gt;, &lt;code&gt;genealogy&lt;/code&gt;, &lt;code&gt;face-recognition&lt;/code&gt;, &lt;code&gt;privacy&lt;/code&gt;, &lt;code&gt;local-first&lt;/code&gt;, &lt;code&gt;opencv&lt;/code&gt;, &lt;code&gt;ollama&lt;/code&gt;, &lt;code&gt;gemma&lt;/code&gt;, &lt;code&gt;python&lt;/code&gt;, &lt;code&gt;hacktoberfest&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;One step only you can do:&lt;/strong&gt; upload &lt;code&gt;docs/branding/social-preview-1280x640.png&lt;/code&gt; in the repo's Settings under "Social preview". GitHub has no way for me to do that. Without it, shared links show a generic card instead of the banner.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;On your side:&lt;/strong&gt; for DEV, replace the images with your uploads from the &lt;code&gt;docs/&lt;/code&gt; files, and swap in the banner as the cover image. The Gemma pass is still running (about 45 minutes more). When it's done, tell me and I'll check that it finished cleanly. I'll also refresh the numbers in the post, so you can publish.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;need draft wiki markup for tool's help space page on wikitree  Space:PhotoTag&lt;br&gt;
keep it clear and easy to understand &lt;br&gt;
all demo photos uploaded&lt;br&gt;
logo and dev banner uploaded&lt;br&gt;
&lt;a href="https://www.wikitree.com/index.php?title=Special:BrowsePhotos&amp;amp;l=57375489" rel="noopener noreferrer"&gt;https://www.wikitree.com/index.php?title=Special:BrowsePhotos&amp;amp;l=57375489&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;g2g post link for space page {{G2GLink|2098697}}&lt;/p&gt;

&lt;p&gt;need draft title and text of g2g post (WYSIWYG format)&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Both drafts are written. The Space page markup is in &lt;a href="docs/wikitree/space-PhotoTag.wiki.txt"&gt;space-PhotoTag.wiki.txt&lt;/a&gt; and the G2G post is in &lt;a href="docs/wikitree/g2g-post.txt"&gt;g2g-post.txt&lt;/a&gt;. Neither is committed or published.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Space:PhotoTag page.&lt;/strong&gt; I checked the live page. It exists but is empty, and it already lists your uploads. The markup uses these exact file names:&lt;br&gt;
- &lt;code&gt;PhotoTag.png&lt;/code&gt;: the logo, at the top right.&lt;br&gt;
- &lt;code&gt;PhotoTag-1.jpg&lt;/code&gt;: name faces.&lt;br&gt;
- &lt;code&gt;PhotoTag-2.jpg&lt;/code&gt;: named people.&lt;br&gt;
- &lt;code&gt;PhotoTag-4.jpg&lt;/code&gt;: find a person.&lt;br&gt;
- &lt;code&gt;PhotoTag-3.jpg&lt;/code&gt;: the group photo.&lt;br&gt;
- &lt;code&gt;PhotoTag.jpg&lt;/code&gt;: the export preview.&lt;/p&gt;

&lt;p&gt;The page has these sections: intro, a "testers welcome" line with &lt;code&gt;{{G2GLink|2098697}}&lt;/code&gt;, what it does, privacy (what is shared and what is not), what you need, how to use it, screenshots, tips, limits, source and credits, and a G2G link at the end. The sample caption is wrapped in &lt;code&gt;&amp;lt;nowiki&amp;gt;&lt;/code&gt; so it shows literally.&lt;/p&gt;

&lt;p&gt;I only found 6 of your 7 uploads, because the gallery showed 5 plus the logo. The seventh is probably the DEV banner, whose file name I couldn't see. It's left off the page. If you want it on the page, tell me the file name.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;G2G post.&lt;/strong&gt; Your thread already has the title "New tool: WikiTree PhotoTag, testers needed", and I think it works well. The file has two alternatives if you want to change it. Here's the body to paste in, with the bold and bullets applied afterwards:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Do you have boxes of family photos on your computer that would make wonderful additions to profiles, but many of the people in them are still living?&lt;/p&gt;

&lt;p&gt;I built a free community tool to help with exactly that, and I'm looking for testers.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;g2g post should be html so i can copy/paste&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;I'll make it an HTML page with a one-click "Copy" button that puts the formatted post on your clipboard, so you can paste it straight into G2G's editor. It also has the raw HTML if you want to paste that into a source view. I'm also rewording the sample caption so G2G doesn't turn it into a broken profile link.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The G2G post is now an HTML page: &lt;a href="docs/wikitree/g2g-post.html"&gt;g2g-post.html&lt;/a&gt;. I checked that it renders correctly. I replaced the old text file.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How to use it:&lt;/strong&gt;&lt;br&gt;
1. Open the file in your own web browser (Chrome or Safari). Double-click it, or use "Reveal in Finder".&lt;br&gt;
2. Click &lt;strong&gt;Copy title&lt;/strong&gt; and paste it into the G2G title box.&lt;br&gt;
3. Click &lt;strong&gt;Copy post body&lt;/strong&gt;, then paste into G2G's editor. The bold text, numbered list, bullets and links come across.&lt;br&gt;
4. If a browser blocks the button, it highlights the post for you instead. Press Cmd+C and paste.&lt;/p&gt;

&lt;p&gt;At the bottom there's a &lt;strong&gt;Raw HTML&lt;/strong&gt; section you can open. Paste that into an editor's source view if the normal paste doesn't keep the formatting.&lt;/p&gt;

&lt;p&gt;I reworded one thing: the sample caption now says "for example, 'Mary Smith with Ann Jones, 1955'" instead of showing the &lt;code&gt;[[Smith-123|Mary Smith]]&lt;/code&gt; link markup. G2G could turn that markup into a broken profile link. The Space page keeps the real markup, safely wrapped in &lt;code&gt;&amp;lt;nowiki&amp;gt;&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The post still has one placeholder: "Write-up: [DEV post link, once published]". Replace it after the DEV post goes live. If you'd like, I'll check on the Gemma pass now, so you know when it's safe to publish.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;check gemma&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The Gemma pass is still running, with 211 photos left. That's about 10 minutes.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Now&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Photos with an era guess&lt;/td&gt;
&lt;td&gt;1,959&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Photos with readable writing&lt;/td&gt;
&lt;td&gt;245 (172 of them have faces, up from 56)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The era guesses are mostly the 1970s–2000s (about 1,750 of 1,959). Only about 30 photos were guessed as 1930s or earlier. That fits a family album where most photos are modern.&lt;/p&gt;

&lt;p&gt;I sampled the writing on photos with faces. It's real handwriting and print, for example:&lt;br&gt;
- a name and birth details ("Jeremiah Alan Sweeny Dec. 21, 1997 7 Lbs. 13 Oz.")&lt;br&gt;
- a ship's name ("Costa Atlantica")&lt;br&gt;
- a park's donation sign&lt;/p&gt;

&lt;p&gt;It also read several printed date stamps, such as "MAY 64" and "89 8 13".&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;An idea you may like, for the date narrowing:&lt;/strong&gt; those date stamps are exactly the year information that was missing, since only 8 of your 2,470 scans had a year before. Right now the tool saves the writing but doesn't turn "MAY 64" into 1964. I can add that. It would be a small change, and it would give more photos a real date instead of a guess. Do you want it?&lt;/p&gt;

&lt;p&gt;I'll check again once it finishes. Before you publish, the post's Gemma claim will be true as written. The only things to update will be the final counts. Tell me when you're ready and I'll refresh them.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;yes. check gemma again in 15 minutes&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Done, and the labeling page is back up at &lt;code&gt;http://&lt;span class="agent-session-redacted"&gt;[REDACTED]&lt;/span&gt;:8765&lt;/code&gt;. It took about 30 seconds to start because the Gemma pass was writing to the database at the same time.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Date stamps are working.&lt;/strong&gt; 45 photos that had no date now have a real year, taken from stamps like "NOV 66" and "MAY 64". That brings you from 8 dated photos to 53, and these years count as strong dating, not as guesses. It only reads short stamps. Longer text is ignored on purpose, because a year in a sentence is as likely to be a birth or a wedding as the date the photo was taken. It never overwrites an existing date. I added two tests (25 pass).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Gemma:&lt;/strong&gt; I've set a background check that reports in up to 15 minutes, or sooner if the pass finishes, and I'll tell you what it finds. At the last check there were 211 photos left, so it should finish within that window.&lt;/p&gt;

&lt;p&gt;Nothing from the date-stamp change is committed or pushed yet. I'll include it in the next commit, and I can do that once Gemma finishes.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;&amp;lt;task-notification&amp;gt;&lt;br&gt;
&amp;lt;task-id&amp;gt;bbmqbga1t&amp;lt;/task-id&amp;gt;&lt;br&gt;
&amp;lt;tool-use-id&amp;gt;toolu_01TyvZPYPbea9ajWtJcKBLaL&amp;lt;/tool-use-id&amp;gt;&lt;br&gt;
&amp;lt;output-file&amp;gt;/private/tmp/claude-501/-Users-azurerobinson-code-Photo-Identifier-Tool/7d115469-52ca-4477-bd62-e37385261d08/tasks/bbmqbga1t.output&amp;lt;/output-file&amp;gt;&lt;br&gt;
&amp;lt;status&amp;gt;completed&amp;lt;/status&amp;gt;&lt;br&gt;
&amp;lt;summary&amp;gt;Background command "Wait up to 15 minutes for the Gemma pass, then report" completed (exit code 0)&amp;lt;/summary&amp;gt;&lt;br&gt;
&amp;lt;/task-notification&amp;gt;&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The Gemma pass finished at 12:59 and nothing is missing. The date-stamp change is committed and pushed (&lt;code&gt;291aa5c&lt;/code&gt;, 25 tests pass).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Final Gemma results:&lt;/strong&gt;&lt;br&gt;
- Every photo has been read: 2,170 photos have an era guess.&lt;br&gt;
- 257 photos have readable writing, and 184 of those have faces.&lt;br&gt;
- 50 photos got an exact year from a print date stamp, so 58 photos now have a real date (up from 8).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Edits for the DEV post, since you're updating it by hand.&lt;/strong&gt; I haven't touched the draft:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Where&lt;/th&gt;
&lt;th&gt;Change&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;"Narrowing by date"&lt;/td&gt;
&lt;td&gt;"only 8 of my 2,470 scans carry a year" → add that Gemma's read of printed date stamps like "MAY 64" raised it to &lt;strong&gt;58&lt;/strong&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"Gemma reads the writing"&lt;/td&gt;
&lt;td&gt;add: &lt;strong&gt;257&lt;/strong&gt; photos had readable writing (184 with faces), and 50 of them were date stamps that became exact photo years&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Lazy Payoff, "What I got"&lt;/td&gt;
&lt;td&gt;"81 photos covering 15 relatives, &lt;strong&gt;739&lt;/strong&gt; stay private" → &lt;strong&gt;767&lt;/strong&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Lazy Payoff, "name the other ~2,800"&lt;/td&gt;
&lt;td&gt;now about &lt;strong&gt;2,750&lt;/strong&gt; faces are unnamed&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Code section&lt;/td&gt;
&lt;td&gt;23 tests → &lt;strong&gt;25&lt;/strong&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The "81 clicks, about 2,500 faces, about 50 minutes" figure stays as is. The tool now shows you've named 2,568 faces in total, and 17 minutes of timed naming since the timer was added. Those aren't part of the 50-minute figure.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Left uncommitted on purpose:&lt;/strong&gt; &lt;code&gt;docs/wikitree/&lt;/code&gt;, with the Space page markup and the G2G post. They're your drafts for WikiTree, not repo content. I can add the help-page markup to the repo if you'd like it kept there.&lt;/p&gt;

&lt;p&gt;The DEV post is ready to publish once you've made those edits. After you do, remember to replace "[DEV post link, once published]" in the G2G post.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;update dev draft post&lt;br&gt;
create new markup for space page and new g2g post content to cover the change&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;The DEV draft is updated, and I kept your image uploads. Now the new Space page markup and G2G post.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The DEV draft, the Space page markup and the G2G post are all updated.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;DEV draft:&lt;/strong&gt; your uploaded images are kept. Changes:&lt;br&gt;
- "Narrowing by date" now covers the print date stamps (8 → 58 dated photos).&lt;br&gt;
- "Gemma reads the writing" has the final numbers: all 2,470 photos read, 257 with writing, 184 of them with faces.&lt;br&gt;
- Lazy Payoff: 81 photos of 15 relatives, 767 stay private, about 2,750 faces unnamed.&lt;br&gt;
- Code section: 25 tests.&lt;br&gt;
- The Gemma prize paragraph now mentions date stamps and the era guess.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Space page:&lt;/strong&gt; &lt;a href="docs/wikitree/space-PhotoTag.wiki.txt"&gt;space-PhotoTag.wiki.txt&lt;/a&gt;&lt;br&gt;
- &lt;strong&gt;New "Smarter suggestions" section:&lt;/strong&gt; it explains in plain words that dates, writing on scans and date stamps, folder and file names, and "older generations first" only change the order of suggestions, so you always decide.&lt;br&gt;
- &lt;strong&gt;Optional Gemma step:&lt;/strong&gt; a line for &lt;code&gt;phototag read-text --all&lt;/code&gt;, with a note that reading a whole album takes a couple of hours in the background and isn't required.&lt;br&gt;
- &lt;strong&gt;Test results:&lt;/strong&gt; the "in one test" line now says 81 photos of 15 deceased relatives were ready to upload.&lt;br&gt;
- &lt;strong&gt;Screenshot tip:&lt;/strong&gt; &lt;code&gt;phototag label --blur&lt;/code&gt;, for anyone sharing a screenshot (for example in G2G) without showing living people.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;G2G post:&lt;/strong&gt; &lt;a href="docs/wikitree/g2g-post.html"&gt;g2g-post.html&lt;/a&gt;. Same copy buttons as before. It gains a "Smarter suggestions (optional)" paragraph and the new test result. There's also a new question for testers: "Do the suggestions help? Tell me if they point you at the wrong person." The title is unchanged.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Still to do on your side:&lt;/strong&gt;&lt;br&gt;
- Replace "[DEV post link, once published]" in the G2G post once you publish.&lt;/p&gt;

&lt;p&gt;[…shortened]&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;a section is missing from my dev post:&lt;/p&gt;
&lt;h2 id="agent-session-1077-80-my-agent-session"&gt;My Agent Session&lt;/h2&gt;

&lt;p&gt;&amp;lt;!-- Optional, but judges love it.  Save your session with DevRelay and embed it with the agent_session tag (see the challenge page), or link to it. --&amp;gt;&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;
  &lt;/div&gt;

  &lt;div class="agent-session-footer"&gt;
    &lt;span class="agent-session-meta"&gt;
        81 of 81 messages
    &lt;/span&gt;
  &lt;/div&gt;
&lt;/div&gt;



&lt;h2&gt;
  
  
  The Lazy Payoff
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Without the tool:&lt;/strong&gt; tag 5,353 faces one at a time, and check each against 4,131 people in the tree.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;With it:&lt;/strong&gt; 81 group decisions covered about &lt;strong&gt;2,500 faces in about 50 minutes&lt;/strong&gt;, roughly 31 faces per click. Setting aside strangers (classmates in school photos, for example) is one more click.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What I got:&lt;/strong&gt; photos of deceased relatives, sorted into WikiTree-ID folders with captions, ready for upload. At the time of writing, 81 photos covering 15 deceased relatives are exportable, while 767 stay private because they include someone treated as possibly living.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What I did not do:&lt;/strong&gt; name the other ~2,750 faces. They are mostly small groups and one-offs, and anything unnamed simply stays private. That is the point of the design: an unfinished job is still a safe one.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Best Use of Gemma:&lt;/strong&gt; Gemma 3 (4B) runs locally through Ollama. It reads handwritten and printed captions and date stamps on the scans (257 photos had writing, 50 gave an exact year), and it makes a rough guess at the decade of every photo, which narrows who can appear in it. Those names and dates feed the suggestions. I'm not using it for the face matching itself, which is what the OpenCV models are for.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>python</category>
      <category>privacy</category>
    </item>
    <item>
      <title>Recipe Roots: a shoebox of recipe cards, a WikiTree profile, and a cookbook your family will fight over</title>
      <dc:creator>Robinson-27225</dc:creator>
      <pubDate>Wed, 14 Oct 2026 02:12:16 +0000</pubDate>
      <link>https://dev.to/robinson27225/recipe-roots-a-shoebox-of-recipe-cards-a-wikitree-profile-and-a-cookbook-your-family-will-fight-52mo</link>
      <guid>https://dev.to/robinson27225/recipe-roots-a-shoebox-of-recipe-cards-a-wikitree-profile-and-a-cookbook-your-family-will-fight-52mo</guid>
      <description>&lt;p&gt;This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/p&gt;

&lt;p&gt;What I Built&lt;br&gt;
Every family has the box. Index cards in three kinds of handwriting, a church cookbook with a cracked spine, a photocopy of a photocopy of someone's mother's pound cake. Everybody agrees it should be "put together someday," and then nobody does, because putting it together means deciding what goes where, who made what, and who even was Aunt Edie.&lt;/p&gt;

&lt;p&gt;Recipe Roots is the "someday" part. You give it your public WikiTree ID and a folder of scans. You get back a print-ready 6×9 in cookbook where every recipe is filed under the relative who cooked it, and every cook gets a page that shows where they were born, who their parents were, what was happening in kitchens during their life, and anything you've written about them.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftqovb2tfg3t3fmji0g2r.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftqovb2tfg3t3fmji0g2r.png" alt=" " width="800" height="1415"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;That is my own family: the tree is mine on WikiTree (public profiles of people who have died), plus seven handwritten recipe cards from my relatives, scanned and attached to the person who cooked each one. My mother is in it too: she is living, so she is only there because she agreed to be named, and her page shows her name, her photo, her parents and only the decade she was born in, never a birth date or a place.&lt;/p&gt;

&lt;p&gt;What it does&lt;/p&gt;

&lt;p&gt;Reads the tree. Ancestors, plus great-aunts and great-uncles and other relatives, from WikiTree's public API. Living people are never included unless you say their family agreed.&lt;br&gt;
Finds the migrations. Born in County Cork, died in Boston: that's a chapter. Plain deterministic code decides this from birth/death/children's places and dates, and it states the move as an honest window ("sometime between 1909 and 1930"), not a guess at the day.&lt;br&gt;
Writes the chapters. An open-weight model (Gemma 3 4B, on my laptop) writes about the food of that place and time. More on why that restriction matters below.&lt;br&gt;
Takes your scans. Photos or PDFs of recipe cards. Each is assigned to a specific person by file name or folder. Gemma's vision reads it into ingredients and directions, or, for handwriting, you can print the scan itself as the page.&lt;br&gt;
Builds the Family Kitchen. A page per cook, oldest first. Their recipes sorted into categories. A contents page with real page numbers and live links to WikiTree profiles. Indices by category and by cook.&lt;br&gt;
Sets the original beside yours. If a public-domain cookbook from their era has the same dish, the book quotes it word for word on the facing page: "How they wrote it, in 1916."&lt;br&gt;
Who would pay&lt;/p&gt;

&lt;p&gt;My hypothesis, not a result: people who want a gift (a keepsake for a reunion or a holiday, for parents or grandparents) and family historians who've already done the genealogy and have a box they can't finish. Someone could also run it as a service: scan the box, produce the book, send it to print. The value is the one nobody does well: loose, uncategorized pages become a finished book, with each recipe beside the person who made it. It has the shape of something you'd order several copies of.&lt;/p&gt;

&lt;p&gt;Demo&lt;br&gt;
Site: reciperoots.tech&lt;br&gt;
Demo book, built from my own tree (PDF, 52 pages): reciperoots.tech/demo-book.pdf&lt;br&gt;
A fully fictional demo, no real family: reciperoots.tech/demo-fictional.pdf&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fts8ut91l5ttr23tylfn0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fts8ut91l5ttr23tylfn0.png" alt=" " width="800" height="1200"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fiwwmaxeuv0ss9uszh0os.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fiwwmaxeuv0ss9uszh0os.png" alt=" " width="800" height="1200"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fntyfgexivx7nrrc29drg.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fntyfgexivx7nrrc29drg.png" alt=" " width="800" height="1200"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg3n1c4e9lu7q4re3s2ar.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg3n1c4e9lu7q4re3s2ar.png" alt=" " width="800" height="1200"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Code&lt;br&gt;
&lt;/p&gt;
&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/robinson27225" rel="noopener noreferrer"&gt;
        robinson27225
      &lt;/a&gt; / &lt;a href="https://github.com/robinson27225/recipe-roots" rel="noopener noreferrer"&gt;
        recipe-roots
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Turn a WikiTree ancestry and scanned family recipes into a print-ready cookbook: a page per cook, recipes by category, and how each dish was written in its own era. Gemma + Postgres/pgvector, runs locally.
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/robinson27225/recipe-roots/docs/images/logo.png"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Frobinson27225%2Frecipe-roots%2FHEAD%2Fdocs%2Fimages%2Flogo.png" alt="Recipe Roots logo" width="140"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Recipe Roots&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;Write-up on DEV: &lt;a href="https://dev.to/robinson27225/recipe-roots-a-shoebox-of-recipe-cards-a-wikitree-profile-and-a-cookbook-your-family-will-fight-52mo" rel="nofollow"&gt;Recipe Roots: a shoebox of recipe cards, a WikiTree profile, and a cookbook your family will fight over&lt;/a&gt;. Site: &lt;a href="https://reciperoots.tech" rel="nofollow noopener noreferrer"&gt;reciperoots.tech&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Turn a public &lt;a href="https://www.wikitree.com" rel="nofollow noopener noreferrer"&gt;WikiTree&lt;/a&gt; ancestry into a print-ready family cookbook (6×9 in PDF): the places your ancestors came from, the food history of their eras, traditional dishes from those kitchens, and &lt;strong&gt;your own family's recipes, scanned from the originals&lt;/strong&gt;.&lt;/p&gt;
&lt;div class="highlight highlight-source-shell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;brew install ollama &lt;span class="pl-k"&gt;&amp;amp;&amp;amp;&lt;/span&gt; ollama pull gemma3:4b      &lt;span class="pl-c"&gt;&lt;span class="pl-c"&gt;#&lt;/span&gt; any OpenAI-compatible open-weight model works&lt;/span&gt;
node src/cli.js Wilder-1 --title &lt;span class="pl-s"&gt;&lt;span class="pl-pds"&gt;"&lt;/span&gt;The Wilder Table&lt;span class="pl-pds"&gt;"&lt;/span&gt;&lt;/span&gt; --dedication &lt;span class="pl-s"&gt;&lt;span class="pl-pds"&gt;"&lt;/span&gt;For Grandma&lt;span class="pl-pds"&gt;"&lt;/span&gt;&lt;/span&gt;
&lt;span class="pl-c"&gt;&lt;span class="pl-c"&gt;#&lt;/span&gt; -&amp;gt; out/Wilder-1/book.pdf&lt;/span&gt;&lt;/pre&gt;

&lt;/div&gt;
&lt;p&gt;Needs Node 20+ and Chrome/Chromium for the PDF (&lt;code&gt;CHROME_PATH&lt;/code&gt; if not found). The only dependency is &lt;code&gt;pg&lt;/code&gt;, and only used with a database
Demo books (PDFs in &lt;a href="https://github.com/robinson27225/recipe-roots/docs" rel="noopener noreferrer"&gt;docs/&lt;/a&gt;): &lt;a href="https://github.com/robinson27225/recipe-roots/docs/demo-book.pdf" rel="noopener noreferrer"&gt;one built from the author's own WikiTree tree&lt;/a&gt;, with seven scanned handwritten recipes from relatives (one of them a living relative…&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/robinson27225/recipe-roots" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;MIT licensed. Node 20+, Chrome for the PDF, and Ollama; the only dependency is pg, and only if you use a database. 125 tests.&lt;/p&gt;

&lt;p&gt;brew install ollama &amp;amp;&amp;amp; ollama pull gemma3:4b&lt;br&gt;
node src/cli.js ancestors Your-1234 --relatives          # find IDs for people to assign recipes to&lt;br&gt;
node src/cli.js scan ./my-scans --tree Your-1234&lt;br&gt;
node src/cli.js review family-recipes.json                # proofread the model's reading&lt;br&gt;
node src/cli.js Your-1234 --family-recipes family-recipes.json --family-people people.json&lt;br&gt;
The book above: node src/cli.js Robinson-27225 --family-recipes examples/robinson-demo/family-recipes.json --title "The Robinson Table" --wikitree-photos --me Robinson-27225 --honor-code-signed --confirm-permission (those last flags are my statement that I uploaded those photos). Fictional demo (no real people): node examples/fictional-demo.mjs.&lt;/p&gt;

&lt;p&gt;How I Built It&lt;br&gt;
I built Recipe Roots with Claude Code, and the models inside the product are open-weight: Gemma 3 4B writes the chapter prose and restructures old receipts, Gemma 3's vision reads the scans, and EmbeddingGemma (768 dimensions) powers search. All of it runs locally through Ollama.&lt;/p&gt;

&lt;p&gt;The model never writes about people&lt;br&gt;
My first version let Gemma write the chapter introductions with everything: names, relationships, dates. It was fluent and confidently wrong. It invented a great-grandfather who "went ahead to join them." A year-checking guard couldn't catch that, because there was no wrong year to catch, just a relationship that did not exist. In a family book that's the worst possible failure: it looks like research and it's fiction.&lt;/p&gt;

&lt;p&gt;So I split the job:&lt;/p&gt;

&lt;p&gt;Code writes every fact about a person: names, kinship, places, years, straight from the tree. A cook's life sentence is assembled from fields ("Born in 1904 in Belle Plaine, Benton, to Carl Budd and Lovey Jane Hoover. Died in 1966 in Aberdeen, Grays Harbor, aged 62.") and age appears only when both dates are full.&lt;br&gt;
The model only writes about food. It's never shown a name, a relation or a birth date, so it has nothing to build a biography from. It sees the region, the era, a kitchen-history note and the dish.&lt;br&gt;
A guard checks every draft: years not in the supplied facts, kinship words, "would have" speculation, invented motives, meta commentary, any person's name. A failed section gets one retry with the reasons, then falls back to a template individually, and the book's "About" page says honestly which parts the model wrote.&lt;br&gt;
I'd rather ship a plainer paragraph than a lovely invented one. The same rule shapes the family pages: anything about a person that isn't in your own notes comes from the tree, not from the model.&lt;/p&gt;

&lt;p&gt;Period receipts: "how they wrote it"&lt;br&gt;
I wanted the book to show a great-grandmother's apple pie next to what an apple pie was in her day. Gutenberg and the Internet Archive have a lot of old American cookbooks, and anything published in the US before 1926 is in the public domain, so the book can quote them: Fannie Farmer (1906), The Virginia Housewife (1860), Eliza Leslie's Philadelphia Directions for Cookery (1840), The Cookery Blue Book of San Francisco (1891), and 15 others, 19 books in all.&lt;/p&gt;

&lt;p&gt;The matching is deliberately strict and has no model in it. A receipt counts only if its title is the same dish with at most a few harmless extra words, so "Marrowfat Beans Baked" is not "Boston Baked Beans" and "Johnny Cake" is not "Skillet Cornbread". No match means no quotation. Heritage and era choose the books: a Pennsylvania line is offered the Pennsylvania books, nearest in time first, and a WikiTree heritage sticker on a profile can offer a heritage cookbook where the community has stated one. It is never inferred from a place name.&lt;/p&gt;

&lt;p&gt;Old receipts are single paragraphs, so under each quotation the book sets them as ingredients-then-method. Where the receipt already lists ingredients, that's plain code. Where it's prose, Gemma restructures it and then a checker holds it to the original: every quantity and ingredient must be in the text, no temperature may be added, no steps repeated. A receipt that fails just shows the original. On eight real 1907 receipts, five came through and three were correctly withheld. What the original leaves out ("the original gives no oven temperature or baking time") is said in one sentence, not filled in.&lt;/p&gt;

&lt;p&gt;Scans, polished&lt;br&gt;
Assign each card to anyone in the tree, including the great-aunt who actually made the thing.&lt;br&gt;
Transcription reads the card twice, the second pass only for numbers, and flags any disagreement, because "tsp" read as "tbsp" ruins a recipe. Until you set "reviewed": true, the page says Transcribed by AI, check amounts.&lt;br&gt;
recipe-roots review writes a sheet with each scan next to what the model read and the doubtful amounts called out.&lt;br&gt;
Handwriting is hard for a 4B model, so there's --page-only: the scan is the page, no model involved. That's usually the right choice for Grandma's own hand.&lt;br&gt;
A small engineering thing I enjoyed: page numbers in a PDF&lt;br&gt;
Chrome's print-to-PDF has no target-counter(), so a contents page can't know page numbers. The fix: put an invisible one-pixel [[marker]] at every section, render, read each marker's page back out of the PDF (PDFKit on macOS, pdftotext elsewhere), render again with real numbers, and repeat until they stop moving. Usually that's two passes. If the PDF can't be read, the contents prints as a plain list, never with wrong numbers.&lt;/p&gt;

&lt;p&gt;Tiger Cloud: hybrid search over every receipt&lt;br&gt;
I loaded every receipt from the 19 public-domain cookbooks, 18,971 of them, into a free Tiger Cloud service (PostgreSQL 18.6) with an EmbeddingGemma vector each, a BM25 index from pg_textsearch and pgvector. It is optional: strict title matching works without a database, and the demo books above were built without it. With --similar-period on, one SQL query ranks the receipts by keyword and by meaning, only within the books that fit the cook's era and region, and fuses the two lists with Reciprocal Rank Fusion (Tiger's own hybrid-search recipe). The 56-dish pantry is embedded there too.&lt;/p&gt;

&lt;p&gt;What hybrid search is for: strict title matching finds the same dish; it can't tell that "Shoofly Pie" is "SHOO-FLY PIE" or that "Spaetzle" is "SPAETZLEN OR SPATZEN". The vector side finds them, a spelling check accepts them as the same dish, and where there's no same dish a receipt is offered as "A similar dish… not the same dish" only if the keyword and vector rankings both put it near the top and it shares the dish's own words. My first, looser rule offered "Turkey Stuffing (Swedish Style)" for Swedish meatballs, so I tightened it.&lt;/p&gt;

&lt;p&gt;The honest numbers: of the 56 pantry dishes, strict matching found 27; hybrid search added 2 spelling variants and 1 similar dish and, correctly, nothing for the other 26 (mostly continental dishes American cookbooks of the time don't print). That is a modest gain, and it's measured, not hoped. A gotcha worth knowing if you use pg_textsearch: inside a CTE or window function the BM25 query fails with "no BM25 index found" unless you name the index with to_bm25query(query, 'index_name'). With 19,000 rows the search is exact, not an ANN index, on purpose. Tiger's certificate comes from a private CA that Node doesn't trust, so I connected with libpq-style TLS (encrypted, not certificate-verified); the README shows how to pin the CA instead.&lt;/p&gt;

&lt;p&gt;What doesn't work yet&lt;br&gt;
The recipes in the chapters aren't your family's, and the book says so. They're written from standard published versions and haven't been kitchen-tested. Cook one before you gift it.&lt;br&gt;
Handwriting transcription is a draft, not a result. Typed and handwriting-font cards read correctly. Real handwriting did not. On the first card Gemma 3 4B read ½ cup cocoa as "2 c." and 2 T butter as "2 t", split one card into two recipes and filed ingredients under directions (I fixed the last two in code). On six more cards it kept dropping the "1" before a fraction (2 1/4 became "2/4", 1 1/4 became "1/4", 1/3 became "3/3"), read "Preheat" as "Present", turned "All chopped fine" into "All Chopped Flounder", invented a method ("Mix all ingredients together") for a card that had none, and missed margin notes like "350°". Only the waffles came through almost clean. The number cross-check flagged most of the wrong amounts, which is why every AI transcription is marked until a person reviews it (I corrected these against the scans and proofread them), why review highlights doubtful lines beside the scan, and why --page-only exists: for handwriting you care about, print the scan itself.&lt;br&gt;
The guard can't verify meaning. It blocks invented years, kin and names, but a model can still connect dots loosely. Read the book before it prints.&lt;br&gt;
28 of 56 dishes have a period receipt, and there is no Midwestern cookbook on Gutenberg that I could find.&lt;br&gt;
Permission. WikiTree photos and recipes belong to their contributors, and this book is meant to be sold. Using them is gated and fails closed: it needs your WikiTree ID, a statement that you've signed the Honor Code (Wiki Genealogists have) and an explicit confirmation, and it writes a ledger of whose work is in the book. WikiTree's API doesn't expose membership level, so the Honor Code check is an attestation the tool can't verify. Your own photos and recipes are never gated.&lt;br&gt;
Structuring a receipt into a different layout doesn't, by itself, get around copyright. That's why the sources are public domain.&lt;br&gt;
No print-on-demand specs yet (bleed, spine, cover): the PDF is a 6×9 interior.&lt;br&gt;
What I'd do next: front-and-back recipe cards, a Midwestern and a German period source, a real print-on-demand cover, a WikiTree login so profiles you're trusted on (not just public ones) can be used, and someone other than me using it. If you've got a box, I'd like to hear what's in it.&lt;/p&gt;

&lt;p&gt;Why Does Open Innovation Matter?&lt;br&gt;
Private by construction. Scans of a family's recipe cards, and their handwriting, are exactly what many people won't upload to a hosted API. The vision model runs on the laptop; nothing leaves it.&lt;br&gt;
Costs nothing per book. A tree-and-scans book needs a few dozen model calls. At local-inference cost, the margin on a gift book is the price of printing, not the price of tokens. That is what makes a side hustle of it.&lt;br&gt;
Controllable. Because the model is local and swappable I could tune the whole system around its weaknesses: restrict it to food, guard its output, fall back per section. Any OpenAI-compatible endpoint works; --model swaps it. A closed API that changed behaviour under me would have broken a guard I had calibrated to this model's failures.&lt;br&gt;
No lock-in to a pricing change on something I'd want to resell.&lt;br&gt;
Open data under it, too. The people come from WikiTree's open API and the period recipes from public-domain books on Project Gutenberg and the Internet Archive. A product built on open models and open sources can be handed to a family, or run as a service, without anyone's terms of service in the middle.&lt;br&gt;
My Agent Session&lt;br&gt;
I built Recipe Roots with Claude Code. These are the highlights of that session, with secrets and private details removed:&lt;/p&gt;


&lt;div class="ltag-agent-session"&gt;
  &lt;div class="agent-session-header"&gt;
    
    &lt;span class="agent-session-tool-icon-badge" title="Claude Code"&gt;
&lt;/span&gt;
    &lt;span class="agent-session-title"&gt;Recipe Roots: building a family cookbook with open models (highlights)&lt;/span&gt;
  &lt;/div&gt;

  &lt;div class="agent-session-scroll"&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;take into consideration = Utilize Partner Categories: Try to integrate technologies that unlock extra cash prizes, such as registering your project on a free .tech domain or using pgvector for your genealogy database search.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;yes, i am Robinson-27225&lt;br&gt;
users would definitely need permission to use a recipe page or image  and are reselling the book&lt;br&gt;
if the person using the app hasn't signed the honor code (indicated by membership level of wiki genealogist) they should not be able to use images or recipes. &lt;br&gt;
for the most part the tool should come back with recipes that are good examples for the time period/heritage of their ancestral lines with the ability to add in actual personal family photos and recipes.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The gate you described is built, but it can't be fully enforced by the tool. The Honor Code status you asked me to check isn't available to it. All 53 tests pass, and I ran the real command line on your tree for the default, personal-photo and gated paths.&lt;/p&gt;
&lt;h2 id="agent-session-880-2-why-the-gate-cant-verify-the-honor-code"&gt;Why the gate can't verify the Honor Code&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What WikiTree says:&lt;/strong&gt; the Help:Membership page confirms there are three levels. Signing the Honor Code is exactly what makes someone a Wiki Genealogist, so your rule fits their model.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What I found:&lt;/strong&gt; your profile page shows "Honor Code Signatory, Signed 25 Apr 2018" and the Wiki Genealogist badge. None of that is in the API. The profile fields, a guessed set of field names, and the profile categories all lack it, and the page can't be read by a script.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What the tool does:&lt;/strong&gt; it &lt;strong&gt;fails closed&lt;/strong&gt;. Using WikiTree photos or other members' recipes needs all three of:

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;--me &amp;lt;ID&amp;gt;&lt;/code&gt;, checked against WikiTree as a real member account,&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;--honor-code-signed&lt;/code&gt;, an attestation the tool can't confirm,&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;--confirm-permission&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Limit of the guardrail:&lt;/strong&gt; it stops honest users by default, but anyone can pass the flags, or paste a recipe into their own file. Real enforcement needs WikiTree to add an API field for membership level or a login flow. You're well placed to ask them.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="agent-session-880-2-whats-gated-and-what-isnt"&gt;What's gated and what isn't&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Gated:&lt;/strong&gt; WikiTree photos, and cookbook recipes written by other members. If you contributed a recipe yourself, it counts as your own work.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Never gated:&lt;/strong&gt; WikiTree facts (names, dates, places), the curated dishes, your own recipes and scans, and your own photos (&lt;code&gt;--my-photos&lt;/code&gt;, named like &lt;code&gt;Smith-1234.jpg&lt;/code&gt; and credited "family collection").&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ledger:&lt;/strong&gt; when any third-party item is used, the build writes &lt;code&gt;permissions.md&lt;/code&gt;, listing each photo and recipe with its owner and source. It records what needs permission. It is not the permission itself.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="agent-session-880-2-heritage-matched-recipes-by-default"&gt;Heritage-matched recipes by default&lt;/h2&gt;

&lt;p&gt;Your tree kept getting apple pie, cornmeal mush and pot pie, because the pantry was thin for the Mid-Atlantic, South, Midwest and Germany. I added 20 traditional dishes for those lines, so the pantry is now 56 dishes in 18 regions. Your book now returns four different corridors, with dishes like scrapple, apple butter, rouladen and chicken and noodles. Chapters also avoid repeating the same corridor. &lt;strong&gt;The dishes are written from standard published versions and checked for plausibility, but I haven't cooked them.&lt;/strong&gt; The README and About page say so, and the About page no longer claims "hand-checked".&lt;/p&gt;
&lt;h2 id="agent-session-880-2-testing-and-cleanup"&gt;Testing and cleanup&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What I ran:&lt;/strong&gt; the default build (no WikiTree content), a personal-photo build, and a gated build. For the gated build, I passed the attestation flags myself purely to exercise the code path. They are not your attestation, and nothing was sold.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cleanup:&lt;/strong&gt; I deleted those outputs, since they contained other members' photos and recipes.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="agent-session-880-2-whats-left"&gt;What's left&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;A living relative's recipes:&lt;/strong&gt; your two cookbook recipes are still excluded because she's living, unless she agrees to be named.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;DEV post:&lt;/strong&gt; I can draft it for you to review, and I won't publish it without your yes. It's due Oct 18, 11:59 PM PDT.&lt;/li&gt;
&lt;/ul&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;so the AI aspect doesn't look at the heritage of the starting profile and use that as a prompt to find recipes that are free available on the web?&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Correct, it doesn't. Today the heritage drives a &lt;strong&gt;deterministic&lt;/strong&gt; step (the migration finder) and a &lt;strong&gt;curated pantry&lt;/strong&gt;. The model only writes the food passages, and the vector search only ranks dishes already in the pantry. Nothing searches the web, and that was partly deliberate: random web recipes are usually copyrighted blog text, which clashes with your resale concern. A web step is still a good idea if it's limited to sources with clear licences. I'll check that such sources can be reached by a script before I promise anything.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;i think one of the values of this will be users wanting to scan their personal family recipes, allowing them to have a polished final product instead of loose uncategorized pages. being able to attribute a recipe to an ancestor and show them in their historical context and their personal family story will be powerful&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;You're describing the product's centre of gravity, and I agree with it. It also shows where today's build falls short of it. A family recipe currently gets a page with its ancestor's name, a story box and a fan chart, but:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;They're loose:&lt;/strong&gt; recipes appear in file order, with no categories, no table of contents, and no index by cook.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No one is introduced:&lt;/strong&gt; the ancestor never gets a page of their own, with their life, their place and their world.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No context:&lt;/strong&gt; nothing connects Grandma's potato salad to how that dish was written in her day.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Before designing the fix, I'll check one technical thing: whether Chrome's PDF engine can print real page numbers in a table of contents and indices. That determines whether they can be real contents pages.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;run it against tiger cloud&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;It works against Tiger Cloud. &lt;code&gt;ingest&lt;/code&gt; loaded all 56 dishes, and a real-tree build (Wilder-1, &lt;code&gt;--offline&lt;/code&gt;, no PDF) ran its semantic search there.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Server:&lt;/strong&gt; the service runs PostgreSQL 18.6 with pgvector 0.8.7 and TimescaleDB 2.30.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dish picks:&lt;/strong&gt; the build wrote chapters with picks like Sourdough Pancakes and Sauerbraten, which differ from the plain picker's. I haven't compared them with the local database, because the local &lt;code&gt;.pgdata&lt;/code&gt; wasn't copied over.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Connection catch:&lt;/strong&gt; a plain &lt;code&gt;sslmode=require&lt;/code&gt; URL fails with &lt;code&gt;self-signed certificate in certificate chain&lt;/code&gt;. Tiger signs with a private CA that Node doesn't trust, and system roots fail in &lt;code&gt;psql&lt;/code&gt; too. Adding &lt;code&gt;&amp;amp;uselibpqcompat=true&lt;/code&gt; made it connect. That means the connection is encrypted but the server certificate isn't verified. To verify it, download the CA certificate from the Tiger console and use &lt;code&gt;sslmode=verify-full&amp;amp;sslrootcert=/path&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Docs:&lt;/strong&gt; the README now carries both options. I rewrote the post's Tiger paragraph to say what ran and name the TLS caveat. The README change is committed.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I didn't write the connection string to any file, and a search of &lt;code&gt;README.md&lt;/code&gt;, &lt;code&gt;docs&lt;/code&gt; and &lt;code&gt;src&lt;/code&gt; finds no copy of the password. It's in this chat, though, so reset it in the Tiger console under the service's settings. I'd do that now, since the test is finished. If you want me to run anything else against the database, I'll need the new one.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Your real card turned up three bugs and a model weakness. I've fixed the bugs, and the weakness is now documented. Test suite: 106 tests pass.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why &lt;code&gt;scan&lt;/code&gt; failed on your second run:&lt;/strong&gt; it re-scanned its own output. The first run saved a copy to &lt;code&gt;~/recipe-test/scans/&lt;/code&gt;, and the second run found it, couldn't tell whose it was, and stopped. It now skips that folder.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How Gemma did on your handwritten pudding card:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Cocoa:&lt;/strong&gt; it read "½ cup cocoa" as "2 c.". That's a fourfold error, the kind that ruins a pudding.&lt;br&gt;
- &lt;strong&gt;Butter:&lt;/strong&gt; it read "2 T butter" as "2 t".&lt;br&gt;
- &lt;strong&gt;Recipe split:&lt;/strong&gt; it split the one card into two recipes. They're now merged by title.&lt;br&gt;
- &lt;strong&gt;Misfiled lines:&lt;/strong&gt; it put some ingredients under directions. A rule now moves lines like "2 1/2 cups whole milk" back, and I removed the stray tick-mark commas. The merged result still isn't in card order.&lt;br&gt;
- &lt;strong&gt;What the checks caught:&lt;/strong&gt; the number cross-check flagged "2" and "1/2", which points at the cocoa, but it missed the butter. The review sheet now highlights lines containing a flagged number.&lt;/p&gt;

&lt;p&gt;For handwriting, use &lt;code&gt;--page-only&lt;/code&gt;, which prints the scan itself as the page. I've written this up honestly in the README and the post's limitations. It's also useful material for the first-user section.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;To redo your test&lt;/strong&gt;, clear the stale output first (the old file has the duplicate recipe), then re-run from the project folder:&lt;/p&gt;

&lt;pre class="highlight bash"&gt;&lt;code&gt;&lt;span class="nb"&gt;rm&lt;/span&gt; &lt;span class="nt"&gt;-rf&lt;/span&gt; ~/recipe-test/scans ~/recipe-test/family-recipes.json
&lt;/code&gt;&lt;/pre&gt;

&lt;pre class="highlight bash"&gt;&lt;code&gt;&lt;span class="nb"&gt;cd&lt;/span&gt; ~/code/recipe-roots &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; node src/cli.js scan ~/recipe-test &lt;span class="nt"&gt;--tree&lt;/span&gt; Robinson-27225 &lt;span class="nt"&gt;--contributor&lt;/span&gt; &lt;span class="s2"&gt;"Your Name"&lt;/span&gt; &lt;span class="nt"&gt;--recipes-file&lt;/span&gt; ~/recipe-test/family-recipes.json
&lt;/code&gt;&lt;/pre&gt;

&lt;p&gt;Then run &lt;code&gt;review&lt;/code&gt; and open &lt;code&gt;review.html&lt;/code&gt; to see the highlights. After that, try the same card with &lt;code&gt;--page-only&lt;/code&gt; for comparison. Tell me which looks better in the finished book.&lt;/p&gt;

&lt;p&gt;The fix is committed locally as &lt;code&gt;0db6781&lt;/code&gt; and not pushed. Run &lt;code&gt;git push&lt;/code&gt; from &lt;code&gt;~/code/recipe-roots&lt;/code&gt;, along with the earlier license commit.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;update dev post with first-user story&lt;/p&gt;

&lt;p&gt;It is powerful to see a handwritten family recipe and the information about that ancestor together on a page, in a book that can be handed down to nieces and nephews. Seeing snippets of what was happening during that time for that ancestor helps me understand what their everyday life may have been like and what they may have made for their family&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The post's first-user section is written, and the placeholder is gone. The edits are in &lt;code&gt;docs/dev-post.md&lt;/code&gt;, which is still local and unpublished.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What I wrote:&lt;/strong&gt;&lt;br&gt;
- It's called "My first user (me, so far)", and it says plainly that nobody has paid. That's honest, and the challenge rewards honesty about this.&lt;br&gt;
- It tells the real story of your pudding card. Gemma read ½ cup cocoa as "2 c." and 2 T butter as "2 t", split the card into two recipes, and misfiled some ingredients. You caught the amounts in &lt;code&gt;review.html&lt;/code&gt; and put the ingredients back in card order. Then you checked Leona's page against what you know of her and it was right.&lt;br&gt;
- Your paragraph is a blockquote, as you wrote it. The point I draw from it: the product is the handwritten card next to who she was and what her kitchen looked like, in something a niece or nephew can hold. The generic regional chapters are the backdrop.&lt;br&gt;
- It ends by saying the next test is putting a finished book in front of a relative who isn't me. I left a line saying you'll report what they said once you have.&lt;/p&gt;

&lt;p&gt;Two small things to check in your voice:&lt;br&gt;
- The section says "two of those bugs were mine to fix and I fixed them". Those fixes were made in this session. Change the wording if you'd rather credit the process.&lt;br&gt;
- It says the card is "filed under" your grandmother, because I don't know whether she wrote it. Edit that if you do.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Left to do before the Oct 18 deadline:&lt;/strong&gt;&lt;br&gt;
- Decide which partner categories to claim and delete the bracket note under "Partner categories". Gemma is solid. Tiger Data is now tested on a real service, though the use is light. &lt;code&gt;.tech&lt;/code&gt; is only claimable if you register a domain.&lt;br&gt;
- Optionally embed your DevRelay session. The challenge invites it, and I can help save one.&lt;br&gt;
- Replace the bracketed draft note at the top, and upload the five images from &lt;code&gt;docs/images/&lt;/code&gt; when you paste it into DEV, since DEV needs them uploaded rather than linked by path.&lt;/p&gt;

&lt;p&gt;Send me any edits, or tell me when you want a final read-through for tone and length.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;The hybrid search now runs on Tiger Cloud, and all of it is pushed (&lt;code&gt;44ac57a&lt;/code&gt;). HTTPS on the domain is still waiting on GitHub.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;On Tiger Cloud:&lt;/strong&gt;&lt;br&gt;
- &lt;strong&gt;Receipts loaded:&lt;/strong&gt; all 18,971 receipts from the 19 cookbooks are in the service, each with an embedding and a BM25 index (&lt;code&gt;pg_textsearch&lt;/code&gt; 1.5.1). The 56-dish pantry is there too.&lt;br&gt;
- &lt;strong&gt;Full build:&lt;/strong&gt; a complete build of your tree with &lt;code&gt;--similar-period&lt;/code&gt; ran against it, using BM25 plus vectors.&lt;br&gt;
- &lt;strong&gt;Gotcha found:&lt;/strong&gt; BM25 queries failed inside the fused search with "no BM25 index found". Naming the index explicitly in the query fixed it. The README and the post both record this.&lt;br&gt;
- &lt;strong&gt;Same results as local:&lt;/strong&gt; the matches matched the local test, which used Postgres's built-in full-text search. Of the 56 pantry dishes, strict matching finds 27. Hybrid search adds 2 spelling variants (Spätzle and Shoofly Pie) and 1 similar dish (Chicken and Noodles), and finds nothing for the other 26. That's a modest gain.&lt;br&gt;
- &lt;strong&gt;Stricter rule:&lt;/strong&gt; a probe with family-style names showed my "similar" rule matched a bare receipt titled "DINNER" for "Roast Beef Dinner". Similar matches now need the dish's own words, never just a mealtime word. The test suite is at 113 passing tests.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Post and docs:&lt;/strong&gt; I rewrote the Tiger section of &lt;code&gt;docs/dev-post.md&lt;/code&gt; around the real results. It leads with what the database does for the book, includes those numbers, and doesn't oversell. The README is updated with the Tiger run, the &lt;code&gt;~/.pgpass&lt;/code&gt; approach, and the deprecation notice node-postgres prints for it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Domain:&lt;/strong&gt; &lt;code&gt;http://reciperoots.tech&lt;/code&gt; returns HTTP 200, and &lt;code&gt;www&lt;/code&gt; redirects to it. HTTPS isn't available yet. GitHub hasn't issued the certificate, and I re-saved the domain to nudge it. This usually clears within an hour of correct DNS. When "Enforce HTTPS" appears in the repo's Pages settings, tick it. Tell me once it does and I'll check that &lt;code&gt;https://reciperoots.tech&lt;/code&gt; loads.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Cleanup:&lt;/strong&gt; I stopped the local test Postgres. Ollama is still running. Your &lt;code&gt;~/.pgpass&lt;/code&gt; still holds the Tiger password. That's fine, but delete the file when you're finished.&lt;/p&gt;

&lt;p&gt;Next: you still need to claim categories (Gemma, Tiger Data, &lt;code&gt;.tech&lt;/code&gt; look supportable now). The first-user paragraph is in. Do you want a full read-through of the post, or shall I add the live domain link first?&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;
  &lt;/div&gt;

  &lt;div class="agent-session-footer"&gt;
    &lt;span class="agent-session-meta"&gt;
        13 of 13 messages
    &lt;/span&gt;
  &lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;My First User&lt;br&gt;
The first person to use it was me, on my own family, so I can't claim a paying customer. Nobody has paid. What I can report is what happened when someone who actually cares about the people in the book used it.&lt;/p&gt;

&lt;p&gt;I took one handwritten recipe card, a chocolate pudding filed under my grandmother, Leona Budd, and ran it through the whole pipeline. It did not go smoothly, and the failures are the useful part. Gemma read "½ cup cocoa" as "2 c.", and "2 T butter" as "2 t". It split the one card into two recipes. It filed some ingredients under the directions. I fixed the split and the misfiled lines in the tool. The cocoa and the butter I caught the way a person would, with review.html open: the scan on one side, the transcription on the other, the doubtful lines highlighted. I corrected them, put the ingredients back into the order on the card, and marked it reviewed. Then I read the page with the facts about her on it, checked them against what I know of her, and they were right. Then I ran six more cards from my relatives through the same steps, including two of my mother's; how that went is under "What doesn't work yet" above.&lt;/p&gt;

&lt;p&gt;What I took from it is not about the model. It's what the page does once it's right. In my words:&lt;/p&gt;

&lt;p&gt;It is powerful to see a handwritten family recipe and the information about that ancestor together, in a book that can be handed down to nieces and nephews. Seeing snippets of what was happening during that time for that ancestor helps me understand what their everyday life may have been like and what they may have made for their family.&lt;/p&gt;

&lt;p&gt;That is the product. Not the chapters of generic regional dishes (those are the backdrop), but the card in her own hand, next to who she was and what her kitchen looked like, in something a niece can hold.&lt;/p&gt;

&lt;p&gt;The next test is the one that counts: put a finished book in front of a relative who isn't me and see whether they ask for another copy. I haven't done that yet, and I'll say here what they said when I have.&lt;/p&gt;

&lt;p&gt;Prize Categories&lt;br&gt;
Best Use of Gemma. Gemma 3 4B writes the chapter prose and restructures old receipts, Gemma 3's vision reads the scans, and EmbeddingGemma powers recipe search. All three run locally through Ollama, and the product is designed around what a 4B model can and can't be trusted with.&lt;br&gt;
Best Use of Tiger Data. A free Tiger Cloud service holds all 18,971 period receipts with a BM25 index (pg_textsearch) and pgvector, and the book's "similar dish" and spelling-variant matching run as one fused hybrid-search query (details above, including the measured numbers and the BM25 gotcha).&lt;br&gt;
Best Use of a .tech Domain. reciperoots.tech is the project's home: a one-page site on GitHub Pages with both demo books, the code and the screenshots, claimed through the Hacktoberfest offer.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
