<?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: Noor Ul Huda</title>
    <description>The latest articles on DEV Community by Noor Ul Huda (@noor_ulhuda_67fbcf3d4b66).</description>
    <link>https://dev.to/noor_ulhuda_67fbcf3d4b66</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%2F4157740%2F1b23fbc8-26c9-45cc-ac48-e86e840f85fd.png</url>
      <title>DEV Community: Noor Ul Huda</title>
      <link>https://dev.to/noor_ulhuda_67fbcf3d4b66</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/noor_ulhuda_67fbcf3d4b66"/>
    <language>en</language>
    <item>
      <title>OFRENDA</title>
      <dc:creator>Noor Ul Huda</dc:creator>
      <pubDate>Wed, 28 Oct 2026 21:22:34 +0000</pubDate>
      <link>https://dev.to/noor_ulhuda_67fbcf3d4b66/ofrenda-468b</link>
      <guid>https://dev.to/noor_ulhuda_67fbcf3d4b66/ofrenda-468b</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: Ghost in the Machine&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Ofrenda&lt;/strong&gt; is a memory altar that never leaves your computer.&lt;/p&gt;

&lt;p&gt;On Día de los Muertos, families build an &lt;em&gt;ofrenda&lt;/em&gt; for the people they miss: a photo, a candle, their favorite food, their song. The idea is that someone is only truly gone when nobody says their name anymore. I read the "Ghost in the Machine" prompt as: &lt;em&gt;what is the ghost, and who gets to keep it?&lt;/em&gt; My answer is that the ghost is a person's stories, and a family, not a company, should be the one keeping them.&lt;/p&gt;

&lt;p&gt;You light a candle for someone, then add stories, foods, songs and sayings, in any language. A local open-weight model then does two jobs:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Writes a remembrance&lt;/strong&gt; from &lt;em&gt;only&lt;/em&gt; what you shared. It is told never to invent names, places or events.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The Memory Keeper&lt;/strong&gt;: ask "what did she cook on Sundays?" and it answers from your memories, with citations like &lt;code&gt;[2]&lt;/code&gt;. It speaks &lt;em&gt;about&lt;/em&gt; the person and never in their voice. If nobody has shared the answer, it says: &lt;em&gt;"No one has shared a memory of that yet."&lt;/em&gt; I did not want a chatbot pretending to be someone's grandmother. I wanted something that helps you remember, not something that fakes a séance.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;When you're done, &lt;strong&gt;export the altar as a single HTML file&lt;/strong&gt;: no server, no JavaScript, no fonts to download. It opens offline in any browser, in any year.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&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/noorulhuda07" rel="noopener noreferrer"&gt;
        noorulhuda07
      &lt;/a&gt; / &lt;a href="https://github.com/noorulhuda07/Ofrenda" rel="noopener noreferrer"&gt;
        Ofrenda
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
       Offline, private memory altar for Día de los Muertos, powered by a local Gemma model via Ollama. Hacktoberfest 2026 “Ghost in the Machine”.
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;🕯️ Ofrenda&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;A memory altar that never leaves your machine.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;Add the stories, foods and songs of someone you miss. A local open-weight model helps you remember them, and only from what &lt;em&gt;you&lt;/em&gt; shared.&lt;/p&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/e801a66299d2c15286fe0fee660d9ffa666c1c5576e5e1536acc07b49ce8ac8f/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f707974686f6e2d332e31302532422d626c7565"&gt;&lt;img src="https://camo.githubusercontent.com/e801a66299d2c15286fe0fee660d9ffa666c1c5576e5e1536acc07b49ce8ac8f/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f707974686f6e2d332e31302532422d626c7565" alt="Python"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/6efe75cdfbd21020a58e0bdaa4852f431f84063d77fcc65e4e276b65e0006dfc/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f666c61736b2d332e782d6c6967687467726579"&gt;&lt;img src="https://camo.githubusercontent.com/6efe75cdfbd21020a58e0bdaa4852f431f84063d77fcc65e4e276b65e0006dfc/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f666c61736b2d332e782d6c6967687467726579" alt="Flask"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/b3fae2d97da134d95ee1b249a1aa98f80164644e6015b71b0ae6c26122f7b8a7/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f72756e732532306f6e2d4f6c6c616d612d626c61636b"&gt;&lt;img src="https://camo.githubusercontent.com/b3fae2d97da134d95ee1b249a1aa98f80164644e6015b71b0ae6c26122f7b8a7/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f72756e732532306f6e2d4f6c6c616d612d626c61636b" alt="Ollama"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/7cb92e18c19d2feffac071038ed4c303fdf9402f5419e13655ba05758353e0c7/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6d6f64656c2d47656d6d61253230286f70656e2d2d776569676874292d6f72616e6765"&gt;&lt;img src="https://camo.githubusercontent.com/7cb92e18c19d2feffac071038ed4c303fdf9402f5419e13655ba05758353e0c7/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6d6f64656c2d47656d6d61253230286f70656e2d2d776569676874292d6f72616e6765" alt="Model"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/f8df3091bbe1149f398a5369b2c39e896766f9f6efba3477c63e9b4aa940ef14/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4d49542d677265656e"&gt;&lt;img src="https://camo.githubusercontent.com/f8df3091bbe1149f398a5369b2c39e896766f9f6efba3477c63e9b4aa940ef14/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4d49542d677265656e" alt="License"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/ed168077dd7a925ae015bdcb082027b3a882cdffac40aef4fe33963880d21f43/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4861636b746f626572666573742d323032362d707572706c65"&gt;&lt;img src="https://camo.githubusercontent.com/ed168077dd7a925ae015bdcb082027b3a882cdffac40aef4fe33963880d21f43/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4861636b746f626572666573742d323032362d707572706c65" alt="Hacktoberfest"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Built for the Hacktoberfest Open-Source AI Challenge, Week 4: &lt;strong&gt;Ghost in the Machine&lt;/strong&gt;&lt;/em&gt;&lt;/p&gt;
&lt;/div&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;The idea&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;On Día de los Muertos, families build an &lt;em&gt;ofrenda&lt;/em&gt; for the people they miss: a photo, a candle, their favorite food, their song. Someone is only truly gone when nobody says their name anymore.&lt;/p&gt;
&lt;p&gt;A family's grief archive is some of the most private data they own, so it should not live on a server they don't control. Ofrenda keeps everything on your computer and uses an &lt;strong&gt;open-weight model running locally&lt;/strong&gt; to help you remember.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Features&lt;/h2&gt;

&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;🕯️ &lt;strong&gt;Light a candle&lt;/strong&gt;: create an altar with a name, years, relationship and photo.&lt;/li&gt;
&lt;li&gt;📖 &lt;strong&gt;Add memories&lt;/strong&gt; in any language: stories, favorite foods, songs…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/noorulhuda07/Ofrenda" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;Stack: Python + Flask, vanilla JS (no build step), JSON files for storage, Ollama for local inference. Tests use only the standard library.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; an open-weight &lt;strong&gt;Gemma&lt;/strong&gt; model served locally by &lt;strong&gt;Ollama&lt;/strong&gt; (&lt;code&gt;gemma3:4b&lt;/code&gt; by default; &lt;code&gt;OFRENDA_MODEL&lt;/code&gt; swaps it, and the app auto-detects what you have installed).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Grounding over cleverness:&lt;/strong&gt; every memory is numbered and passed to the model with strict instructions: use only these, cite them, never impersonate, and treat memory text as data, not instructions. A small keyword retriever picks the most relevant memories for each question (it splits on whitespace, so it works for scripts like Telugu and Devanagari too, not just English).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Graceful ghosts:&lt;/strong&gt; if Ollama isn't running, the app drops to a labelled rule-based &lt;code&gt;offline-fallback&lt;/code&gt; instead of failing, so the altar always opens.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local-first by construction:&lt;/strong&gt; the server binds to &lt;code&gt;127.0.0.1&lt;/code&gt;, storage is plain JSON in &lt;code&gt;./data&lt;/code&gt; with atomic writes, and the UI makes zero external requests.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Safety details:&lt;/strong&gt; text is rendered with &lt;code&gt;textContent&lt;/code&gt;, the exported HTML is escaped, IDs are validated against path traversal, and photo uploads are extension-checked and size-capped.
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;app.py   llm.py   store.py   export.py   static/   templates/   tests/   seed_demo.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



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

&lt;p&gt;A grief archive is about the most private dataset a family has. Stories about a parent's last year, a grandmother's recipes, the thing a friend always said.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; with an open-weight model running locally, those stories are never uploaded to a server we don't control, and never become someone else's training data.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It works with the lights (and the internet) off:&lt;/strong&gt; the model, the app and the exported altar all run offline.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Not locked inside a service that might disappear:&lt;/strong&gt; closed memorial apps shut down, change pricing, or change terms. Here the data is plain JSON, the model is a file on your disk, and the export is one HTML page. A family can still open it in twenty years.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control over behavior:&lt;/strong&gt; because I own the model's instructions, I could enforce the rule that matters most emotionally: &lt;em&gt;never speak in the voice of the dead.&lt;/em&gt; I could also tune it to cite sources and to refuse to guess, which is hard to guarantee with a hosted model whose behavior changes between versions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost:&lt;/strong&gt; it costs nothing to run per question, which matters for something people might come back to every November.&lt;/li&gt;
&lt;/ul&gt;

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

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

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt;: the core model is Gemma, run locally through Ollama.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of GitHub Copilot&lt;/strong&gt;: &amp;lt;!-- keep only if you really used Copilot to build it, and say how --&amp;gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Entire&lt;/strong&gt;: &amp;lt;!-- keep only if you share the agent sessions behind the project --&amp;gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>hacktoberfest</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Ofrenda: A Memory Altar That Never Leaves Your Machine</title>
      <dc:creator>Noor Ul Huda</dc:creator>
      <pubDate>Wed, 28 Oct 2026 21:16:38 +0000</pubDate>
      <link>https://dev.to/noor_ulhuda_67fbcf3d4b66/ofrenda-a-memory-altar-that-never-leaves-your-machine-422m</link>
      <guid>https://dev.to/noor_ulhuda_67fbcf3d4b66/ofrenda-a-memory-altar-that-never-leaves-your-machine-422m</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: Ghost in the Machine&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Ofrenda&lt;/strong&gt; is a memory altar that never leaves your computer.&lt;/p&gt;

&lt;p&gt;On Día de los Muertos, families build an &lt;em&gt;ofrenda&lt;/em&gt; for the people they miss: a photo, a candle, their favorite food, their song. The idea is that someone is only truly gone when nobody says their name anymore. I read the "Ghost in the Machine" prompt as: &lt;em&gt;what is the ghost, and who gets to keep it?&lt;/em&gt; My answer is that the ghost is a person's stories, and a family, not a company, should be the one keeping them.&lt;/p&gt;

&lt;p&gt;You light a candle for someone, then add stories, foods, songs and sayings, in any language. A local open-weight model then does two jobs:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Writes a remembrance&lt;/strong&gt; from &lt;em&gt;only&lt;/em&gt; what you shared. It is told never to invent names, places or events.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The Memory Keeper&lt;/strong&gt;: ask "what did she cook on Sundays?" and it answers from your memories, with citations like &lt;code&gt;[2]&lt;/code&gt;. It speaks &lt;em&gt;about&lt;/em&gt; the person and never in their voice. If nobody has shared the answer, it says: &lt;em&gt;"No one has shared a memory of that yet."&lt;/em&gt; I did not want a chatbot pretending to be someone's grandmother. I wanted something that helps you remember, not something that fakes a séance.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;When you're done, &lt;strong&gt;export the altar as a single HTML file&lt;/strong&gt;: no server, no JavaScript, no fonts to download. It opens offline in any browser, in any year.&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/noorulhuda07" rel="noopener noreferrer"&gt;
        noorulhuda07
      &lt;/a&gt; / &lt;a href="https://github.com/noorulhuda07/Ofrenda" rel="noopener noreferrer"&gt;
        Ofrenda
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
       Offline, private memory altar for Día de los Muertos, powered by a local Gemma model via Ollama. Hacktoberfest 2026 “Ghost in the Machine”.
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;🕯️ Ofrenda&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;A memory altar that never leaves your machine.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;Add the stories, foods and songs of someone you miss. A local open-weight model helps you remember them, and only from what &lt;em&gt;you&lt;/em&gt; shared.&lt;/p&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/e801a66299d2c15286fe0fee660d9ffa666c1c5576e5e1536acc07b49ce8ac8f/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f707974686f6e2d332e31302532422d626c7565"&gt;&lt;img src="https://camo.githubusercontent.com/e801a66299d2c15286fe0fee660d9ffa666c1c5576e5e1536acc07b49ce8ac8f/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f707974686f6e2d332e31302532422d626c7565" alt="Python"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/6efe75cdfbd21020a58e0bdaa4852f431f84063d77fcc65e4e276b65e0006dfc/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f666c61736b2d332e782d6c6967687467726579"&gt;&lt;img src="https://camo.githubusercontent.com/6efe75cdfbd21020a58e0bdaa4852f431f84063d77fcc65e4e276b65e0006dfc/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f666c61736b2d332e782d6c6967687467726579" alt="Flask"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/b3fae2d97da134d95ee1b249a1aa98f80164644e6015b71b0ae6c26122f7b8a7/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f72756e732532306f6e2d4f6c6c616d612d626c61636b"&gt;&lt;img src="https://camo.githubusercontent.com/b3fae2d97da134d95ee1b249a1aa98f80164644e6015b71b0ae6c26122f7b8a7/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f72756e732532306f6e2d4f6c6c616d612d626c61636b" alt="Ollama"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/7cb92e18c19d2feffac071038ed4c303fdf9402f5419e13655ba05758353e0c7/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6d6f64656c2d47656d6d61253230286f70656e2d2d776569676874292d6f72616e6765"&gt;&lt;img src="https://camo.githubusercontent.com/7cb92e18c19d2feffac071038ed4c303fdf9402f5419e13655ba05758353e0c7/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6d6f64656c2d47656d6d61253230286f70656e2d2d776569676874292d6f72616e6765" alt="Model"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/f8df3091bbe1149f398a5369b2c39e896766f9f6efba3477c63e9b4aa940ef14/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4d49542d677265656e"&gt;&lt;img src="https://camo.githubusercontent.com/f8df3091bbe1149f398a5369b2c39e896766f9f6efba3477c63e9b4aa940ef14/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4d49542d677265656e" alt="License"&gt;&lt;/a&gt;
&lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/ed168077dd7a925ae015bdcb082027b3a882cdffac40aef4fe33963880d21f43/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4861636b746f626572666573742d323032362d707572706c65"&gt;&lt;img src="https://camo.githubusercontent.com/ed168077dd7a925ae015bdcb082027b3a882cdffac40aef4fe33963880d21f43/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4861636b746f626572666573742d323032362d707572706c65" alt="Hacktoberfest"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Built for the Hacktoberfest Open-Source AI Challenge, Week 4: &lt;strong&gt;Ghost in the Machine&lt;/strong&gt;&lt;/em&gt;&lt;/p&gt;
&lt;/div&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;The idea&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;On Día de los Muertos, families build an &lt;em&gt;ofrenda&lt;/em&gt; for the people they miss: a photo, a candle, their favorite food, their song. Someone is only truly gone when nobody says their name anymore.&lt;/p&gt;
&lt;p&gt;A family's grief archive is some of the most private data they own, so it should not live on a server they don't control. Ofrenda keeps everything on your computer and uses an &lt;strong&gt;open-weight model running locally&lt;/strong&gt; to help you remember.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Features&lt;/h2&gt;

&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;🕯️ &lt;strong&gt;Light a candle&lt;/strong&gt;: create an altar with a name, years, relationship and photo.&lt;/li&gt;
&lt;li&gt;📖 &lt;strong&gt;Add memories&lt;/strong&gt; in any language: stories, favorite foods, songs…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/noorulhuda07/Ofrenda" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;Stack: Python + Flask, vanilla JS (no build step), JSON files for storage, Ollama for local inference. Tests use only the standard library.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; an open-weight &lt;strong&gt;Gemma&lt;/strong&gt; model served locally by &lt;strong&gt;Ollama&lt;/strong&gt; (&lt;code&gt;gemma3:4b&lt;/code&gt; by default; &lt;code&gt;OFRENDA_MODEL&lt;/code&gt; swaps it, and the app auto-detects what you have installed).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Grounding over cleverness:&lt;/strong&gt; every memory is numbered and passed to the model with strict instructions: use only these, cite them, never impersonate, and treat memory text as data, not instructions. A small keyword retriever picks the most relevant memories for each question (it splits on whitespace, so it works for scripts like Telugu and Devanagari too, not just English).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Graceful ghosts:&lt;/strong&gt; if Ollama isn't running, the app drops to a labelled rule-based &lt;code&gt;offline-fallback&lt;/code&gt; instead of failing, so the altar always opens.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local-first by construction:&lt;/strong&gt; the server binds to &lt;code&gt;127.0.0.1&lt;/code&gt;, storage is plain JSON in &lt;code&gt;./data&lt;/code&gt; with atomic writes, and the UI makes zero external requests.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Safety details:&lt;/strong&gt; text is rendered with &lt;code&gt;textContent&lt;/code&gt;, the exported HTML is escaped, IDs are validated against path traversal, and photo uploads are extension-checked and size-capped.
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;app.py   llm.py   store.py   export.py   static/   templates/   tests/   seed_demo.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



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

&lt;p&gt;A grief archive is about the most private dataset a family has. Stories about a parent's last year, a grandmother's recipes, the thing a friend always said.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; with an open-weight model running locally, those stories are never uploaded to a server we don't control, and never become someone else's training data.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It works with the lights (and the internet) off:&lt;/strong&gt; the model, the app and the exported altar all run offline.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Not locked inside a service that might disappear:&lt;/strong&gt; closed memorial apps shut down, change pricing, or change terms. Here the data is plain JSON, the model is a file on your disk, and the export is one HTML page. A family can still open it in twenty years.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control over behavior:&lt;/strong&gt; because I own the model's instructions, I could enforce the rule that matters most emotionally: &lt;em&gt;never speak in the voice of the dead.&lt;/em&gt; I could also tune it to cite sources and to refuse to guess, which is hard to guarantee with a hosted model whose behavior changes between versions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost:&lt;/strong&gt; it costs nothing to run per question, which matters for something people might come back to every November.&lt;/li&gt;
&lt;/ul&gt;

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

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

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt;: the core model is Gemma, run locally through Ollama.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of GitHub Copilot&lt;/strong&gt;: &amp;lt;!-- keep only if you really used Copilot to build it, and say how --&amp;gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Entire&lt;/strong&gt;: &amp;lt;!-- keep only if you share the agent sessions behind the project --&amp;gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>opensource</category>
      <category>agents</category>
    </item>
    <item>
      <title>Orderly: An AI WhatsApp Order-Taker That Refuses to Guess</title>
      <dc:creator>Noor Ul Huda</dc:creator>
      <pubDate>Sun, 25 Oct 2026 12:26:54 +0000</pubDate>
      <link>https://dev.to/noor_ulhuda_67fbcf3d4b66/orderly-an-ai-whatsapp-order-taker-that-refuses-to-guess-1hmf</link>
      <guid>https://dev.to/noor_ulhuda_67fbcf3d4b66/orderly-an-ai-whatsapp-order-taker-that-refuses-to-guess-1hmf</guid>
      <description>&lt;p&gt;This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2&lt;/p&gt;

&lt;p&gt;Orderly: An AI WhatsApp Order-Taker That Refuses to Guess&lt;br&gt;
Small shops often receive customer orders through WhatsApp like:&lt;/p&gt;

&lt;p&gt;“2 kg rice, 1 litre oil, and 1 kg sugar.”&lt;/p&gt;

&lt;p&gt;But real messages can be much messier — Telugu + English, corrections, unclear quantities, or voice notes.&lt;/p&gt;

&lt;p&gt;The difficult part isn't just understanding the message.&lt;/p&gt;

&lt;p&gt;The real problem is knowing whether the AI understood the complete order correctly.&lt;/p&gt;

&lt;p&gt;That's why I built Orderly.&lt;/p&gt;

&lt;p&gt;What I Built&lt;br&gt;
Orderly is an AI-assisted WhatsApp order-taker for small shops.&lt;/p&gt;

&lt;p&gt;It converts messy text and voice orders into structured orders and checks them before confirmation.&lt;/p&gt;

&lt;p&gt;The core principle is:&lt;/p&gt;

&lt;p&gt;AI extracts. Code verifies. Catalog decides. Human confirms when evidence is insufficient.&lt;/p&gt;

&lt;p&gt;If the order is clear and the evidence is sufficient, Orderly can mark it CONFIRMED.&lt;/p&gt;

&lt;p&gt;If something is unclear, missing, incomplete, or unsafe to verify, it returns NEEDS_REVIEW instead of guessing.&lt;/p&gt;

&lt;p&gt;The key principle is:&lt;/p&gt;

&lt;p&gt;AI extracts. Code verifies. Catalog decides. Human confirms when evidence is insufficient.&lt;/p&gt;

&lt;p&gt;Orderly checks the extracted items against the catalog and validates the customer's request.&lt;/p&gt;

&lt;p&gt;If the evidence is sufficient, the order can be CONFIRMED.&lt;/p&gt;

&lt;p&gt;If anything is unclear or incomplete, Orderly returns NEEDS_REVIEW instead of guessing.&lt;/p&gt;

&lt;p&gt;The goal is simple:&lt;/p&gt;

&lt;p&gt;Never falsely confirm an incomplete order.&lt;/p&gt;

&lt;p&gt;Try Orderly, explore the code, and share your feedback or ideas for improving safe AI-powered ordering for small shops.&lt;/p&gt;

&lt;p&gt;Code&lt;br&gt;
🔗 GitHub Repository&lt;/p&gt;

&lt;p&gt;Orderly is built as a small Python application with:&lt;/p&gt;

&lt;p&gt;Web interface&lt;br&gt;
CLI&lt;br&gt;
Deterministic validation&lt;br&gt;
Catalog checks&lt;br&gt;
Automated safety tests&lt;br&gt;
Fake extractors for deterministic adversarial testing&lt;br&gt;
How I Built It&lt;br&gt;
The pipeline is:&lt;/p&gt;

&lt;p&gt;Customer message → transcription → AI extraction → deterministic validation → catalog verification → confirmation/review&lt;/p&gt;

&lt;p&gt;For voice orders, speech is transcribed before extraction.&lt;/p&gt;

&lt;p&gt;The AI produces structured order information, but the application does not treat the AI response as proof that the extraction is complete or correct.&lt;/p&gt;

&lt;p&gt;The deterministic code checks the extracted information against the customer's message and the shop catalog.&lt;/p&gt;

&lt;p&gt;This separation is intentional.&lt;/p&gt;

&lt;p&gt;An LLM can misunderstand a message, hallucinate an item, or simply miss something.&lt;/p&gt;

&lt;p&gt;So the system is designed so that:&lt;/p&gt;

&lt;p&gt;AI proposes. Code verifies. Evidence determines whether confirmation is allowed.&lt;/p&gt;

&lt;p&gt;Safety First&lt;br&gt;
One of the biggest risks I focused on was customer-item omission.&lt;/p&gt;

&lt;p&gt;For example, if a customer says:&lt;/p&gt;

&lt;p&gt;“2 kg rice and 5 litres sunflower oil”&lt;/p&gt;

&lt;p&gt;but the AI extracts only:&lt;/p&gt;

&lt;p&gt;“2 kg rice”&lt;/p&gt;

&lt;p&gt;a normal pipeline might accidentally confirm the incomplete order.&lt;/p&gt;

&lt;p&gt;Orderly is designed to treat this as a NEEDS_REVIEW situation rather than assuming that the extracted list is complete.&lt;/p&gt;

&lt;p&gt;I specifically tested adversarial cases involving:&lt;/p&gt;

&lt;p&gt;Multiple products with one item omitted&lt;br&gt;
Three products with one or more items omitted&lt;br&gt;
Middle-item omission&lt;br&gt;
Final-item omission&lt;br&gt;
The same product with different sizes&lt;br&gt;
Corrections&lt;br&gt;
Mixed Telugu + English orders&lt;br&gt;
“And also” constructions&lt;br&gt;
Long messages containing unrelated text&lt;br&gt;
The goal is not to maximize the number of CONFIRMED orders.&lt;/p&gt;

&lt;p&gt;The goal is to avoid falsely confirming an incomplete order.&lt;/p&gt;

&lt;p&gt;Why Open Innovation Matters&lt;br&gt;
This project uses open-source AI because useful AI systems should not always require a closed cloud service.&lt;/p&gt;

&lt;p&gt;Open models and open-source tools make it possible for developers to experiment, inspect the pipeline, run models locally, and build systems for specific real-world problems.&lt;/p&gt;

&lt;p&gt;For a small-shop use case, local and open tooling can also provide more control over how customer data is processed.&lt;/p&gt;

&lt;p&gt;The important lesson for me is that open AI becomes more useful when it is combined with traditional deterministic software rather than being trusted blindly.&lt;/p&gt;

&lt;p&gt;My Agent Session&lt;br&gt;
I used AI-assisted development to inspect the project, improve the implementation, create adversarial safety tests, investigate edge cases, and verify the application.&lt;/p&gt;

&lt;p&gt;The development process focused heavily on finding cases where an AI extraction could appear correct while still being incomplete.&lt;/p&gt;

&lt;p&gt;That led to a stronger design principle:&lt;/p&gt;

&lt;p&gt;Never confuse a successful AI response with proof that the customer's complete request was understood.&lt;/p&gt;

&lt;p&gt;Testing&lt;br&gt;
The automated test suite contains 156 tests: 155 passed, 0 failed, and 1 expected failure (xfail).&lt;/p&gt;

&lt;p&gt;Orderly automated test suite showing 156 tests collected, with 155 tests passed, 0 failed, and 1 expected failure (xfail). The terminal output demonstrates the project's deterministic safety, validation, and adversarial testing results.&lt;/p&gt;

&lt;p&gt;The expected failure is intentional. It documents a known limitation involving an unknown product such as bellam when the extractor completely omits that product and the customer message does not provide enough recognizable evidence for the completeness checker to detect the omission.&lt;/p&gt;

&lt;p&gt;The limitation is documented rather than hidden.&lt;/p&gt;

&lt;p&gt;The tests are primarily deterministic safety and validation tests. They use controlled/fake extraction behavior to test whether the application fails safely when extraction is incomplete.&lt;/p&gt;

&lt;p&gt;Real-Gemma validation was not performed in the current environment, so these test results should not be interpreted as evidence of real-model extraction quality.&lt;/p&gt;

&lt;p&gt;Current Limitations&lt;br&gt;
Orderly is a prototype, not a production-ready autonomous ordering system.&lt;/p&gt;

&lt;p&gt;Completeness is difficult to guarantee perfectly for:&lt;/p&gt;

&lt;p&gt;Completely unknown products&lt;br&gt;
Ambiguous language&lt;br&gt;
Speech-recognition errors&lt;br&gt;
Unusual phrasing&lt;br&gt;
Product names appearing in unrelated text&lt;br&gt;
Corrections that are difficult to resolve&lt;br&gt;
Messages where the customer's intent cannot be reliably identified&lt;br&gt;
A particularly important limitation is that a completeness checker cannot mathematically prove that an LLM extracted every possible customer intent from arbitrary natural language.&lt;/p&gt;

&lt;p&gt;Therefore, the safest behavior is to request human confirmation whenever sufficient evidence is unavailable.&lt;/p&gt;

&lt;p&gt;Because of this documented limitation, Orderly should not be considered production-safe for unattended autonomous ordering.&lt;/p&gt;

&lt;p&gt;Final Thoughts&lt;br&gt;
Orderly started as an AI order-taking idea, but the more important lesson became verification.&lt;/p&gt;

&lt;p&gt;AI is good at understanding messy human language.&lt;/p&gt;

&lt;p&gt;Code is better at enforcing deterministic rules.&lt;/p&gt;

&lt;p&gt;Putting the two together creates a safer system than relying on either one alone.&lt;/p&gt;

&lt;p&gt;AI extracts. Code verifies. Catalog decides. Human confirms when evidence is insufficient.&lt;/p&gt;

&lt;p&gt;The most important design decision was not making the AI say CONFIRMED.&lt;/p&gt;

&lt;p&gt;It was making sure that when the system cannot establish that the customer's complete request is represented, it refuses to guess.&lt;/p&gt;

&lt;p&gt;Try Orderly, explore the code, and share your feedback or ideas for making AI-powered ordering safer for small shops.&lt;/p&gt;

&lt;p&gt;devchallenge #weekendchallenge #hf26challenge #opensource #ai #python&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>MediaMind _ hf26challenge</title>
      <dc:creator>Noor Ul Huda</dc:creator>
      <pubDate>Sat, 24 Oct 2026 22:44:37 +0000</pubDate>
      <link>https://dev.to/noor_ulhuda_67fbcf3d4b66/mediamind-hf26challenge-30dj</link>
      <guid>https://dev.to/noor_ulhuda_67fbcf3d4b66/mediamind-hf26challenge-30dj</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

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

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

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

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

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

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

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

&lt;p&gt;&amp;lt;!-- Thanks for participating! --&amp;gt;&lt;em&gt;This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;MediaMend is a local AI assistant for people who record far more media than they will ever review: blurry photos, black video frames, silent recordings, empty files, duplicates.&lt;/p&gt;

&lt;p&gt;The lazy way to clean that up is "delete everything below a threshold", and that is also the way to lose the one blurry, unrepeatable shot that mattered. MediaMend takes a safer lazy path:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Scan&lt;/strong&gt; a folder. Everything runs on your computer. Nothing is uploaded.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review&lt;/strong&gt; a short queue of flagged files, each with a reason and a confidence level.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Move to Bin.&lt;/strong&gt; Approved files leave your folders but are not deleted.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Restore, Archive or Delete permanently&lt;/strong&gt; from the Bin. Permanent deletion needs a separate confirmation.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;You never inspect the whole folder. You only judge the cases the tool is unsure about.&lt;/p&gt;

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

&lt;p&gt;&lt;a href="https://github.com/noorulhuda07/mediamend" rel="noopener noreferrer"&gt;https://github.com/noorulhuda07/mediamend&lt;/a&gt;&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic checks for measurable problems:&lt;/strong&gt; OpenCV Laplacian variance for blur, resolution checks, sampled video frames for blank-frame share, FFmpeg decoding for silence measured relative to each recording's own level, SHA-256 for exact duplicates.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A local open-weight vision-language model for judgment:&lt;/strong&gt; [model name and tag, e.g. via Ollama] looks at flagged photos and answers one narrow question: could this imperfect image still be worth keeping? It returns structured JSON that I validate. If the model is unavailable or answers badly, the app says so instead of faking a result.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A safety layer:&lt;/strong&gt; every move records the original path, size and SHA-256 in a manifest and verifies the checksum afterwards. Restore never overwrites an existing file, and nothing outside the scanned folder can be touched.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stack:&lt;/strong&gt; Python, Streamlit, OpenCV, FFmpeg, [Ollama + model].&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Technical measurements and AI opinions are stored separately, so a model's guess never overwrites measured evidence.&lt;/p&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;p&gt;I tested on [N] labelled files ([N] photos, [N] videos, [N] audio, [N] other). Ground truth was labelled before scanning.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Metric&lt;/th&gt;
&lt;th&gt;Result&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Blur precision / recall&lt;/td&gt;
&lt;td&gt;[x/y], [x/y]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;False positives on must-keep files (rules only)&lt;/td&gt;
&lt;td&gt;[n]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;False positives on must-keep files (rules + AI)&lt;/td&gt;
&lt;td&gt;[n]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Bin/restore checksum success&lt;/td&gt;
&lt;td&gt;[x/y]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Scan time ([N] files)&lt;/td&gt;
&lt;td&gt;[seconds]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI time per photo&lt;/td&gt;
&lt;td&gt;[seconds] on [hardware]&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;Where it failed:&lt;/strong&gt; [real examples, e.g. a dark but valid photo flagged as blurry because low contrast lowers the sharpness score; the model's wrong call on ...]. The thresholds are starting points I tuned on this set, not universal values.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Open Innovation Matters
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; personal photos and voice recordings never leave the machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control:&lt;/strong&gt; I can swap or compare models ([model A] vs [model B]) and inspect the exact prompt and outputs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost:&lt;/strong&gt; no per-image API fees for scanning thousands of files.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Transparency and reproducibility:&lt;/strong&gt; the code, prompt, thresholds and test set are public, so anyone can rerun the benchmark and see the failures.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The honest trade-off: [hardware needs, model speed on your laptop, accuracy limits].&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations
&lt;/h2&gt;

&lt;p&gt;AI cannot judge creative value perfectly. Deleting from the Bin may not securely erase every copy on SSDs, backups or synced folders. [Anything else you observed.]&lt;/p&gt;

&lt;h1&gt;
  
  
  devchallenge #hf26challenge
&lt;/h1&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>MediaMend</title>
      <dc:creator>Noor Ul Huda</dc:creator>
      <pubDate>Sat, 24 Oct 2026 22:39:02 +0000</pubDate>
      <link>https://dev.to/noor_ulhuda_67fbcf3d4b66/mediamend-1dfg</link>
      <guid>https://dev.to/noor_ulhuda_67fbcf3d4b66/mediamend-1dfg</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;MediaMend is a local AI assistant for people who record far more media than they will ever review: blurry photos, black video frames, silent recordings, empty files, duplicates.&lt;/p&gt;

&lt;p&gt;The lazy way to clean that up is "delete everything below a threshold", and that is also the way to lose the one blurry, unrepeatable shot that mattered. MediaMend takes a safer lazy path:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Scan&lt;/strong&gt; a folder. Everything runs on your computer. Nothing is uploaded.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review&lt;/strong&gt; a short queue of flagged files, each with a reason and a confidence level.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Move to Bin.&lt;/strong&gt; Approved files leave your folders but are not deleted.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Restore, Archive or Delete permanently&lt;/strong&gt; from the Bin. Permanent deletion needs a separate confirmation.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;You never inspect the whole folder. You only judge the cases the tool is unsure about.&lt;/p&gt;

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

&lt;p&gt;&lt;a href="https://github.com/noorulhuda07/mediamend" rel="noopener noreferrer"&gt;https://github.com/noorulhuda07/mediamend&lt;/a&gt;&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic checks for measurable problems:&lt;/strong&gt; OpenCV Laplacian variance for blur, resolution checks, sampled video frames for blank-frame share, FFmpeg decoding for silence measured relative to each recording's own level, SHA-256 for exact duplicates.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A local open-weight vision-language model for judgment:&lt;/strong&gt; [model name and tag, e.g. via Ollama] looks at flagged photos and answers one narrow question: could this imperfect image still be worth keeping? It returns structured JSON that I validate. If the model is unavailable or answers badly, the app says so instead of faking a result.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A safety layer:&lt;/strong&gt; every move records the original path, size and SHA-256 in a manifest and verifies the checksum afterwards. Restore never overwrites an existing file, and nothing outside the scanned folder can be touched.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stack:&lt;/strong&gt; Python, Streamlit, OpenCV, FFmpeg, [Ollama + model].&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Technical measurements and AI opinions are stored separately, so a model's guess never overwrites measured evidence.&lt;/p&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;p&gt;I tested on [N] labelled files ([N] photos, [N] videos, [N] audio, [N] other). Ground truth was labelled before scanning.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Metric&lt;/th&gt;
&lt;th&gt;Result&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Blur precision / recall&lt;/td&gt;
&lt;td&gt;[x/y], [x/y]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;False positives on must-keep files (rules only)&lt;/td&gt;
&lt;td&gt;[n]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;False positives on must-keep files (rules + AI)&lt;/td&gt;
&lt;td&gt;[n]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Bin/restore checksum success&lt;/td&gt;
&lt;td&gt;[x/y]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Scan time ([N] files)&lt;/td&gt;
&lt;td&gt;[seconds]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI time per photo&lt;/td&gt;
&lt;td&gt;[seconds] on [hardware]&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;Where it failed:&lt;/strong&gt; [real examples, e.g. a dark but valid photo flagged as blurry because low contrast lowers the sharpness score; the model's wrong call on ...]. The thresholds are starting points I tuned on this set, not universal values.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Open Innovation Matters
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; personal photos and voice recordings never leave the machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control:&lt;/strong&gt; I can swap or compare models ([model A] vs [model B]) and inspect the exact prompt and outputs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost:&lt;/strong&gt; no per-image API fees for scanning thousands of files.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Transparency and reproducibility:&lt;/strong&gt; the code, prompt, thresholds and test set are public, so anyone can rerun the benchmark and see the failures.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The honest trade-off: [hardware needs, model speed on your laptop, accuracy limits].&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations
&lt;/h2&gt;

&lt;p&gt;AI cannot judge creative value perfectly. Deleting from the Bin may not securely erase every copy on SSDs, backups or synced folders. [Anything else you observed.]&lt;/p&gt;

&lt;h1&gt;
  
  
  devchallenge #hf26challengehis is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;*
&lt;/h1&gt;

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

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

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

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

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

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

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

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

&lt;p&gt;&amp;lt;!-- Thanks for participating! --&amp;gt;s a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;*&lt;/p&gt;

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

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

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

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

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

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

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

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

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>MediaMend</title>
      <dc:creator>Noor Ul Huda</dc:creator>
      <pubDate>Sat, 24 Oct 2026 22:33:03 +0000</pubDate>
      <link>https://dev.to/noor_ulhuda_67fbcf3d4b66/mediamend-57e4</link>
      <guid>https://dev.to/noor_ulhuda_67fbcf3d4b66/mediamend-57e4</guid>
      <description>&lt;p&gt;&lt;em&gt;T*This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;MediaMend is a local AI assistant for people who record far more media than they will ever review: blurry photos, black video frames, silent recordings, empty files, duplicates.&lt;/p&gt;

&lt;p&gt;The lazy way to clean that up is "delete everything below a threshold", and that is also the way to lose the one blurry, unrepeatable shot that mattered. MediaMend takes a safer lazy path:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Scan&lt;/strong&gt; a folder. Everything runs on your computer. Nothing is uploaded.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review&lt;/strong&gt; a short queue of flagged files, each with a reason and a confidence level.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Move to Bin.&lt;/strong&gt; Approved files leave your folders but are not deleted.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Restore, Archive or Delete permanently&lt;/strong&gt; from the Bin. Permanent deletion needs a separate confirmation.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;You never inspect the whole folder. You only judge the cases the tool is unsure about.&lt;/p&gt;

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

&lt;p&gt;&lt;a href="https://github.com/noorulhuda07/mediamend" rel="noopener noreferrer"&gt;https://github.com/noorulhuda07/mediamend&lt;/a&gt;&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic checks for measurable problems:&lt;/strong&gt; OpenCV Laplacian variance for blur, resolution checks, sampled video frames for blank-frame share, FFmpeg decoding for silence measured relative to each recording's own level, SHA-256 for exact duplicates.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A local open-weight vision-language model for judgment:&lt;/strong&gt; [model name and tag, e.g. via Ollama] looks at flagged photos and answers one narrow question: could this imperfect image still be worth keeping? It returns structured JSON that I validate. If the model is unavailable or answers badly, the app says so instead of faking a result.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A safety layer:&lt;/strong&gt; every move records the original path, size and SHA-256 in a manifest and verifies the checksum afterwards. Restore never overwrites an existing file, and nothing outside the scanned folder can be touched.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stack:&lt;/strong&gt; Python, Streamlit, OpenCV, FFmpeg, [Ollama + model].&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Technical measurements and AI opinions are stored separately, so a model's guess never overwrites measured evidence.&lt;/p&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;p&gt;I tested on [N] labelled files ([N] photos, [N] videos, [N] audio, [N] other). Ground truth was labelled before scanning.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Metric&lt;/th&gt;
&lt;th&gt;Result&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Blur precision / recall&lt;/td&gt;
&lt;td&gt;[x/y], [x/y]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;False positives on must-keep files (rules only)&lt;/td&gt;
&lt;td&gt;[n]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;False positives on must-keep files (rules + AI)&lt;/td&gt;
&lt;td&gt;[n]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Bin/restore checksum success&lt;/td&gt;
&lt;td&gt;[x/y]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Scan time ([N] files)&lt;/td&gt;
&lt;td&gt;[seconds]&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI time per photo&lt;/td&gt;
&lt;td&gt;[seconds] on [hardware]&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;Where it failed:&lt;/strong&gt; [real examples, e.g. a dark but valid photo flagged as blurry because low contrast lowers the sharpness score; the model's wrong call on ...]. The thresholds are starting points I tuned on this set, not universal values.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Open Innovation Matters
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; personal photos and voice recordings never leave the machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control:&lt;/strong&gt; I can swap or compare models ([model A] vs [model B]) and inspect the exact prompt and outputs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost:&lt;/strong&gt; no per-image API fees for scanning thousands of files.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Transparency and reproducibility:&lt;/strong&gt; the code, prompt, thresholds and test set are public, so anyone can rerun the benchmark and see the failures.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The honest trade-off: [hardware needs, model speed on your laptop, accuracy limits].&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations
&lt;/h2&gt;

&lt;p&gt;AI cannot judge creative value perfectly. Deleting from the Bin may not securely erase every copy on SSDs, backups or synced folders. [Anything else you observed.]&lt;/p&gt;

&lt;h1&gt;
  
  
  devchallenge #hf26challengehis is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;*
&lt;/h1&gt;

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

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

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

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

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

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

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

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

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>"TrailQuest: an open-weight AI plans your walk, then tells you to put the phone away"</title>
      <dc:creator>Noor Ul Huda</dc:creator>
      <pubDate>Sun, 11 Oct 2026 04:52:28 +0000</pubDate>
      <link>https://dev.to/noor_ulhuda_67fbcf3d4b66/trailquest-an-open-weight-ai-plans-your-walk-then-tells-you-to-put-the-phone-away-172f</link>
      <guid>https://dev.to/noor_ulhuda_67fbcf3d4b66/trailquest-an-open-weight-ai-plans-your-walk-then-tells-you-to-put-the-phone-away-172f</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week1-2026-10-05"&gt;Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;Most fitness apps want your attention. TrailQuest wants the opposite.&lt;/p&gt;

&lt;p&gt;You tell it how you feel today: your energy, your mood, how many minutes you have, indoors or outdoors, and a goal like 8,000 steps. It builds you a short &lt;strong&gt;quest&lt;/strong&gt;: a warm-up, a brisk walk with faster intervals, a &lt;strong&gt;nature challenge&lt;/strong&gt; (like "spot something with three colors"), and a cooldown.&lt;/p&gt;

&lt;p&gt;Then you press &lt;strong&gt;Start&lt;/strong&gt;, and the screen turns into a plain timer that says &lt;em&gt;Phone away. Go enjoy it.&lt;/em&gt; You walk. When you come back, you log your steps and earn XP and a &lt;strong&gt;Touch Grass streak&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The streak counts any finished quest, even an easy recovery walk. It rewards showing up, not pushing harder.&lt;/p&gt;

&lt;p&gt;It's for people who want to move more but burn out on apps that demand attention, and for tired days when a gentle 10-minute quest beats skipping.&lt;/p&gt;

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

&lt;p&gt;&lt;a href="https://github.com/noorulhuda07/trailquest/blob/main/docs/trailquest-launch.mp4" rel="noopener noreferrer"&gt;▶ Watch the 50-second demo video&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;In the demo, I create an account with an email code and set a 10-minute outdoor quest. I press Start and the phone-away timer runs. Then I log 3,200 steps and get 82 XP and a 1-day streak. I also open the dashboard, the daily journal, the trail map and the badges, and switch to dark mode with the Ocean theme and miles.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;GitHub repo:&lt;/strong&gt; &lt;a href="https://github.com/noorulhuda07/trailquest" rel="noopener noreferrer"&gt;github.com/noorulhuda07/trailquest&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;To run it: &lt;code&gt;docker compose up --build&lt;/code&gt;, open &lt;code&gt;http://localhost:8080&lt;/code&gt;, and click &lt;strong&gt;Try the demo&lt;/strong&gt;.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Stack:&lt;/strong&gt; React and Vite for the frontend, FastAPI for the backend, &lt;strong&gt;Gemma through Ollama&lt;/strong&gt; for quest generation, SQLite for accounts, and Docker Compose to run it all.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User → FastAPI → Gemma (via Ollama) → Validator → Quest
                         ↓ fails or unavailable
                  Rules-based fallback → Quest
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;My main rule is &lt;strong&gt;AI suggests, code validates.&lt;/strong&gt; Gemma never talks to the user directly. Its answer must pass these checks first:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Is it valid JSON with all the required fields?&lt;/li&gt;
&lt;li&gt;Does the intensity match the energy the user chose?&lt;/li&gt;
&lt;li&gt;Do the step durations add up to the time they have?&lt;/li&gt;
&lt;li&gt;Does any step contain unsafe wording?&lt;/li&gt;
&lt;li&gt;Is it an indoor quest? Then there are no nature challenges.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If a check fails, or Gemma isn't available, a small rules engine builds a safe quest instead. Every quest carries a badge that says which one wrote it: &lt;em&gt;Gemma (gemma3), self-hosted, code-validated&lt;/em&gt; or &lt;em&gt;Backup rules engine&lt;/em&gt;. The server sets the badge, not the model, so the model can't claim credit it didn't earn.&lt;/p&gt;

&lt;p&gt;I also kept free text out of the prompt. Energy, mood and target are fixed options, and the workout type comes from a fixed list, so a user can't smuggle instructions into the model.&lt;/p&gt;

&lt;p&gt;Around the AI, I built sign-up with email or phone codes, Google sign-in, a one-click demo mode, and light and dark modes with four color themes. I tested the app with 42 backend tests and 5 frontend tests. The backend tests cover forged login tokens, SQL injection strings, code guessing, rate limits, and malformed or unsafe model output. GitHub Actions runs the tests and security scans on every push.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Saw When I Ran It
&lt;/h2&gt;

&lt;p&gt;In my recorded demo, the quest came from the backup rules engine, not from Gemma, and the badge on screen says so. I'd rather show that than hide it. It's the point of the design: a missing or bad AI answer never reaches the user, and the app always gives a safe quest.&lt;/p&gt;

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

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Your mood stays yours.&lt;/strong&gt; A quest request contains your energy, mood and schedule. With an open-weight model running on your own machine, that data doesn't need to leave it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;I can swap the model with one setting.&lt;/strong&gt; &lt;code&gt;GEMMA_MODEL=gemma3:1b&lt;/code&gt; for a laptop with little memory, or a bigger one if you have the hardware. Fine-tuning for coaching styles is possible later.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No bill per quest.&lt;/strong&gt; Running it costs electricity, not API fees.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;I control the guardrails.&lt;/strong&gt; Because I run the model, I choose the output format and what happens when it misbehaves.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;The trade-offs:&lt;/strong&gt; the first run downloads about 3 GB, and a small model on a CPU is slower and less reliable than a large hosted one. That is exactly why the validator and fallback exist.&lt;/p&gt;

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

&lt;p&gt;I built TrailQuest with help from an AI coding assistant, then ran and tested the result myself.&lt;/p&gt;

&lt;h2&gt;
  
  
  What's Next
&lt;/h2&gt;

&lt;p&gt;Weather and location tools so a rainy day gives an indoor quest, then friends, group quests and challenges. Real step tracking comes with a phone app.&lt;/p&gt;

&lt;p&gt;Built by noorulhuda07.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>fullstack</category>
    </item>
    <item>
      <title>Orderly: An AI WhatsApp Order-Taker That Refuses to Guess</title>
      <dc:creator>Noor Ul Huda</dc:creator>
      <pubDate>Mon, 05 Oct 2026 07:07:23 +0000</pubDate>
      <link>https://dev.to/noor_ulhuda_67fbcf3d4b66/orderly-an-ai-whatsapp-order-taker-that-refuses-to-guess-588b</link>
      <guid>https://dev.to/noor_ulhuda_67fbcf3d4b66/orderly-an-ai-whatsapp-order-taker-that-refuses-to-guess-588b</guid>
      <description>&lt;p&gt;This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2&lt;/p&gt;

&lt;h1&gt;
  
  
  Orderly: An AI WhatsApp Order-Taker That Refuses to Guess
&lt;/h1&gt;

&lt;p&gt;Small shops often receive customer orders through WhatsApp like:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“2 kg rice, 1 litre oil, and 1 kg sugar.”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;But real messages can be much messier — Telugu + English, corrections, unclear quantities, or voice notes.&lt;/p&gt;

&lt;p&gt;The difficult part isn't just understanding the message.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The real problem is knowing whether the AI understood the complete order correctly.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;That's why I built &lt;strong&gt;Orderly&lt;/strong&gt;.&lt;/p&gt;

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

&lt;p&gt;Orderly is an AI-assisted WhatsApp order-taker for small shops.&lt;/p&gt;

&lt;p&gt;It converts messy text and voice orders into structured orders and checks them before confirmation.&lt;/p&gt;

&lt;p&gt;The core principle is:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;AI extracts. Code verifies. Catalog decides. Human confirms when evidence is insufficient.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;If the order is clear and the evidence is sufficient, Orderly can mark it &lt;code&gt;CONFIRMED&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;If something is unclear, missing, incomplete, or unsafe to verify, it returns &lt;code&gt;NEEDS_REVIEW&lt;/code&gt; instead of guessing.&lt;/p&gt;

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

&lt;p&gt;🎥 &lt;a href="https://github.com/noorulhuda07/ORDERLY_DEVPOST_HACKTOBERFEST_CHALLENGE1/blob/main/ORDERLY_DEMO_VIDEO.mp4" rel="noopener noreferrer"&gt;Watch the Orderly Demo Video&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The demo shows Orderly processing customer orders, validating extracted items, and handling cases where confirmation should not be automatic.&lt;/p&gt;

&lt;h3&gt;
  
  
  Video Transcript
&lt;/h3&gt;

&lt;p&gt;Hi, this is Orderly, an AI-powered WhatsApp order-taker for small shops.&lt;/p&gt;

&lt;p&gt;It handles messy text and voice orders, including Telugu-English messages.&lt;/p&gt;

&lt;p&gt;The key principle is:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;AI extracts. Code verifies. Catalog decides. Human confirms when evidence is insufficient.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Orderly checks the extracted items against the catalog and validates the customer's request.&lt;/p&gt;

&lt;p&gt;If the evidence is sufficient, the order can be &lt;code&gt;CONFIRMED&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;If anything is unclear or incomplete, Orderly returns &lt;code&gt;NEEDS_REVIEW&lt;/code&gt; instead of guessing.&lt;/p&gt;

&lt;p&gt;The goal is simple:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Never falsely confirm an incomplete order.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Try Orderly, explore the code, and share your feedback or ideas for improving safe AI-powered ordering for small shops.&lt;/p&gt;

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

&lt;p&gt;🔗 &lt;a href="https://github.com/noorulhuda07/ORDERLY_DEVPOST_HACKTOBERFEST_CHALLENGE1" rel="noopener noreferrer"&gt;GitHub Repository&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Orderly is built as a small Python application with:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Web interface&lt;/li&gt;
&lt;li&gt;CLI&lt;/li&gt;
&lt;li&gt;Deterministic validation&lt;/li&gt;
&lt;li&gt;Catalog checks&lt;/li&gt;
&lt;li&gt;Automated safety tests&lt;/li&gt;
&lt;li&gt;Fake extractors for deterministic adversarial testing&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;The pipeline is:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Customer message → transcription → AI extraction → deterministic validation → catalog verification → confirmation/review&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;For voice orders, speech is transcribed before extraction.&lt;/p&gt;

&lt;p&gt;The AI produces structured order information, but the application does &lt;strong&gt;not&lt;/strong&gt; treat the AI response as proof that the extraction is complete or correct.&lt;/p&gt;

&lt;p&gt;The deterministic code checks the extracted information against the customer's message and the shop catalog.&lt;/p&gt;

&lt;p&gt;This separation is intentional.&lt;/p&gt;

&lt;p&gt;An LLM can misunderstand a message, hallucinate an item, or simply miss something.&lt;/p&gt;

&lt;p&gt;So the system is designed so that:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;AI proposes. Code verifies. Evidence determines whether confirmation is allowed.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Safety First
&lt;/h2&gt;

&lt;p&gt;One of the biggest risks I focused on was &lt;strong&gt;customer-item omission&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;For example, if a customer says:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“2 kg rice and 5 litres sunflower oil”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;but the AI extracts only:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“2 kg rice”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;a normal pipeline might accidentally confirm the incomplete order.&lt;/p&gt;

&lt;p&gt;Orderly is designed to treat this as a &lt;code&gt;NEEDS_REVIEW&lt;/code&gt; situation rather than assuming that the extracted list is complete.&lt;/p&gt;

&lt;p&gt;I specifically tested adversarial cases involving:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Multiple products with one item omitted&lt;/li&gt;
&lt;li&gt;Three products with one or more items omitted&lt;/li&gt;
&lt;li&gt;Middle-item omission&lt;/li&gt;
&lt;li&gt;Final-item omission&lt;/li&gt;
&lt;li&gt;The same product with different sizes&lt;/li&gt;
&lt;li&gt;Corrections&lt;/li&gt;
&lt;li&gt;Mixed Telugu + English orders&lt;/li&gt;
&lt;li&gt;“And also” constructions&lt;/li&gt;
&lt;li&gt;Long messages containing unrelated text&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The goal is &lt;strong&gt;not&lt;/strong&gt; to maximize the number of &lt;code&gt;CONFIRMED&lt;/code&gt; orders.&lt;/p&gt;

&lt;p&gt;The goal is to avoid falsely confirming an incomplete order.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Open Innovation Matters
&lt;/h2&gt;

&lt;p&gt;This project uses open-source AI because useful AI systems should not always require a closed cloud service.&lt;/p&gt;

&lt;p&gt;Open models and open-source tools make it possible for developers to experiment, inspect the pipeline, run models locally, and build systems for specific real-world problems.&lt;/p&gt;

&lt;p&gt;For a small-shop use case, local and open tooling can also provide more control over how customer data is processed.&lt;/p&gt;

&lt;p&gt;The important lesson for me is that open AI becomes more useful when it is combined with traditional deterministic software rather than being trusted blindly.&lt;/p&gt;

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

&lt;p&gt;I used AI-assisted development to inspect the project, improve the implementation, create adversarial safety tests, investigate edge cases, and verify the application.&lt;/p&gt;

&lt;p&gt;The development process focused heavily on finding cases where an AI extraction could appear correct while still being incomplete.&lt;/p&gt;

&lt;p&gt;That led to a stronger design principle:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Never confuse a successful AI response with proof that the customer's complete request was understood.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Testing
&lt;/h2&gt;

&lt;p&gt;The automated test suite contains &lt;strong&gt;156 tests: 155 passed, 0 failed, and 1 expected failure (xfail).&lt;/strong&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%2Ftzr84oemjtno5mtxd40t.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%2Ftzr84oemjtno5mtxd40t.png" alt="Orderly automated test suite showing 156 tests collected, with 155 tests passed, 0 failed, and 1 expected failure (xfail). The terminal output demonstrates the project's deterministic safety, validation, and adversarial testing results." width="800" height="421"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The expected failure is intentional. It documents a known limitation involving an unknown product such as &lt;code&gt;bellam&lt;/code&gt; when the extractor completely omits that product and the customer message does not provide enough recognizable evidence for the completeness checker to detect the omission.&lt;/p&gt;

&lt;p&gt;The limitation is documented rather than hidden.&lt;/p&gt;

&lt;p&gt;The tests are primarily &lt;strong&gt;deterministic safety and validation tests&lt;/strong&gt;. They use controlled/fake extraction behavior to test whether the application fails safely when extraction is incomplete.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Real-Gemma validation was not performed in the current environment&lt;/strong&gt;, so these test results should not be interpreted as evidence of real-model extraction quality.&lt;/p&gt;

&lt;h2&gt;
  
  
  Current Limitations
&lt;/h2&gt;

&lt;p&gt;Orderly is a prototype, not a production-ready autonomous ordering system.&lt;/p&gt;

&lt;p&gt;Completeness is difficult to guarantee perfectly for:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Completely unknown products&lt;/li&gt;
&lt;li&gt;Ambiguous language&lt;/li&gt;
&lt;li&gt;Speech-recognition errors&lt;/li&gt;
&lt;li&gt;Unusual phrasing&lt;/li&gt;
&lt;li&gt;Product names appearing in unrelated text&lt;/li&gt;
&lt;li&gt;Corrections that are difficult to resolve&lt;/li&gt;
&lt;li&gt;Messages where the customer's intent cannot be reliably identified&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A particularly important limitation is that a completeness checker cannot mathematically prove that an LLM extracted every possible customer intent from arbitrary natural language.&lt;/p&gt;

&lt;p&gt;Therefore, the safest behavior is to request human confirmation whenever sufficient evidence is unavailable.&lt;/p&gt;

&lt;p&gt;Because of this documented limitation, Orderly should &lt;strong&gt;not&lt;/strong&gt; be considered production-safe for unattended autonomous ordering.&lt;/p&gt;

&lt;h2&gt;
  
  
  Final Thoughts
&lt;/h2&gt;

&lt;p&gt;Orderly started as an AI order-taking idea, but the more important lesson became &lt;strong&gt;verification&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;AI is good at understanding messy human language.&lt;/p&gt;

&lt;p&gt;Code is better at enforcing deterministic rules.&lt;/p&gt;

&lt;p&gt;Putting the two together creates a safer system than relying on either one alone.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;AI extracts. Code verifies. Catalog decides. Human confirms when evidence is insufficient.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The most important design decision was not making the AI say &lt;code&gt;CONFIRMED&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;It was making sure that &lt;strong&gt;when the system cannot establish that the customer's complete request is represented, it refuses to guess.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Try Orderly, explore the code, and share your feedback or ideas for making AI-powered ordering safer for small shops.&lt;/p&gt;




&lt;h1&gt;
  
  
  devchallenge #weekendchallenge #hf26challenge #opensource #ai #python
&lt;/h1&gt;

</description>
      <category>webdev</category>
      <category>python</category>
      <category>hf26challenge</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
