<?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: Jed Aureus Gonzales</title>
    <description>The latest articles on DEV Community by Jed Aureus Gonzales (@jedau).</description>
    <link>https://dev.to/jedau</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%2F505927%2F66576a1e-2bc9-4caa-bfe2-3e265c7db096.jpeg</url>
      <title>DEV Community: Jed Aureus Gonzales</title>
      <link>https://dev.to/jedau</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/jedau"/>
    <language>en</language>
    <item>
      <title>Purpoise: Slow and Steady to find your purpose</title>
      <dc:creator>Jed Aureus Gonzales</dc:creator>
      <pubDate>Sat, 24 Oct 2026 11:56:16 +0000</pubDate>
      <link>https://dev.to/jedau/purpoise-slow-and-steady-to-find-your-purpose-3ah2</link>
      <guid>https://dev.to/jedau/purpoise-slow-and-steady-to-find-your-purpose-3ah2</guid>
      <description>&lt;p&gt;This is my submission to the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge, Week 3&lt;/a&gt;. The theme is Slow and Steady.&lt;/p&gt;

&lt;p&gt;You answer one question a week. Ninety seconds, by voice. purpoise does the rest: it reads the signals you already generate, and tells you honestly whether your work carries your purpose, partly carries it, crowds it out, or has nothing to do with it.&lt;/p&gt;

&lt;p&gt;Ikigai is not your job. Most apps about purpose get that wrong, and then quietly turn into career coaching in a kimono.&lt;/p&gt;

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

&lt;p&gt;purpoise keeps two models. One is your purpose. One is your work. A verdict compares them. The purpose model is built with no reference to an employer, a job title, or an industry. Work is inferred from work-channel signals, and only those. Mixing the two is how a job gets mislabelled as a reason for being.&lt;/p&gt;

&lt;p&gt;The week looks like this. You open the app. It asks one question, chosen because a circle is still thin. You answer. The rest of the week it reads what you already produced: a purchases export, bookmarks, a social download, mail, a calendar, commits. It redacts, extracts signals, and files each signal under love, good-at, needs, or paid. A trait that only ever appears at work is hatched and capped until it shows up in a second channel. The verdict says which circles are present, which are missing, and which claim sits near a miss. Every claim opens onto dated evidence. Reject one and the verdict is computed again.&lt;/p&gt;

&lt;p&gt;The person who gets to be lazier is the one who would otherwise take a fifty-question purpose quiz and still not know whether Tuesday's work had anything to do with the answer. The quiz is replaced by one question and a pile of records they already have.&lt;/p&gt;

&lt;p&gt;The client is live at &lt;a href="https://jedau.github.io/purpoise/" rel="noopener noreferrer"&gt;https://jedau.github.io/purpoise/&lt;/a&gt;. That Pages build is the last push to main. The walkthrough below was recorded from the local client, which is ahead of that push.&lt;/p&gt;

&lt;p&gt;The file is &lt;a href="https://github.com/jedau/purpoise/blob/main/docs/demo.mp4" rel="noopener noreferrer"&gt;&lt;code&gt;docs/demo.mp4&lt;/code&gt;&lt;/a&gt;. It is 26 seconds, silent, 1280×800. The blob link works once this file is on main. It is not embedded in this draft, because the draft was written before that push. It is also not the three-minute shot list in &lt;code&gt;docs/DEMO.md&lt;/code&gt;. That list still includes airplane mode, a live redaction, and a twelve-week replay. None of those are in this recording.&lt;/p&gt;

&lt;p&gt;What the recording shows, in order:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The four circles overlap. The middle reads Ikigai, and a purpose sentence marked as a reading.&lt;/li&gt;
&lt;li&gt;A tap on love opens sketching, night walks, and Sunday soup. A second tap puts the middle back.&lt;/li&gt;
&lt;li&gt;A tap on good-at opens letters, and a hatched Tuesday status report marked work only.&lt;/li&gt;
&lt;li&gt;The ledger. Teaching is first, with &lt;code&gt;[teach-1]&lt;/code&gt; and &lt;code&gt;[teach-2]&lt;/code&gt;. Thursday briefings is the paid circle, with &lt;code&gt;[brief-1]&lt;/code&gt; and &lt;code&gt;[brief-2]&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;The verdict opens on "Your work carries it." The middle is Ikigai. Passion, Mission, Profession, and Vocation are all present.&lt;/li&gt;
&lt;li&gt;The page moves down the four questions. Love holds sketching, night walks, and Sunday soup. Good-at holds letters, and a hatched status report marked work only. Needs holds teaching. Paid holds the Thursday briefings.&lt;/li&gt;
&lt;li&gt;Reject on teaching. The card says the verdict will move.&lt;/li&gt;
&lt;li&gt;The verdict again: "Your work carries part of it." The middle is open. Mission and Vocation are gone. Needs says it is missing. Paid still holds the Thursday briefings.&lt;/li&gt;
&lt;li&gt;Consent, as three facts and three groups: bring a file, sign in, request a file.&lt;/li&gt;
&lt;li&gt;The circles again. The needs circle is empty, and the middle says it is open.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  How it is built
&lt;/h2&gt;

&lt;p&gt;Nine stages, in order: source, redact, extract, corroborate, volition, circle map, diagnose, verdict, synthesise. The weekly answer writes labels back into the same file.&lt;/p&gt;

&lt;p&gt;Volition is placement. It decides where a signal lands. It does not decide whether the signal counts. The score is a local TabPFN sidecar when that process is up, and a documented heuristic when it is not. The heuristic is &lt;code&gt;0.5 + 0.2&lt;/code&gt; outside work hours, &lt;code&gt;+ 0.15&lt;/code&gt; on a weekend, &lt;code&gt;+ 0.15&lt;/code&gt; if you started the thread, clamped, and it cannot fall below 0.5. Work hours are 09:00–17:00 UTC. A weekday work-channel vector sent to the sidecar, with the licence key set, came back near 0.29. Stopping the sidecar flags the span and keeps the heuristic. Both paths are tested.&lt;/p&gt;

&lt;p&gt;Text inference is local Ollama. The extract bake-off scored four installed models on four labelled sentences. All four matched every gold kind, so F1 was 1.000 for each. The registry then picked the lower median latency, &lt;code&gt;llama3.2:3b&lt;/code&gt; at 926 ms. &lt;code&gt;qwen3:4b&lt;/code&gt; was 1333 ms, &lt;code&gt;gemma3:4b&lt;/code&gt; 1633 ms, &lt;code&gt;mistral:7b&lt;/code&gt; 2086 ms. Cost per thousand was 0, because the process is on the machine. &lt;code&gt;n&lt;/code&gt; is 4, so the table is indicative. The 500-record replay was not run. A later call to Backboard authenticated and then refused chat, because that free credit is reserved for Memory and RAG. The refusal was not stored. The table stayed the local run.&lt;/p&gt;

&lt;p&gt;The held-out check is a different, smaller thing. Eight rows train a per-subject mode. Two rows are held out. Accuracy on those two is 1.000 against a naive baseline of 0.500. The confidence interval is 0.342 to 1.000. &lt;code&gt;n&lt;/code&gt; is 2, so that row is indicative too. The Tinker API was not called. The delta in &lt;code&gt;docs/EVAL.md&lt;/code&gt; is that local subject mode against a global mode: accuracy 0.500 to 1.000, macro-F1 0.333 to 1.000, cost 0.&lt;/p&gt;

&lt;p&gt;The web client has no on-device model. Mobile is the tier that can redact and extract on the phone, through Flutter Edge AI, and the settings screen says which tier ran. The live agreement against twenty fixtures was not measured. No phone was connected, and the int4 file was not downloaded. The scorer itself is tested: 18 of 20 matching kinds passes, 17 does not.&lt;/p&gt;

&lt;p&gt;The API is Fastify. The worker process stays up and does not start Temporal, so the activities run inside the request. MongoDB and Tiger Data are in the design and were not written during this build. The character file the demo shows is the in-app ledger.&lt;/p&gt;

&lt;h2&gt;
  
  
  What stays on the machine
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Capability&lt;/th&gt;
&lt;th&gt;Where it ran&lt;/th&gt;
&lt;th&gt;Leaves the machine?&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Extracting a signal from a sentence&lt;/td&gt;
&lt;td&gt;Local Ollama&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Choosing the extract model&lt;/td&gt;
&lt;td&gt;Local bake-off, four models&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Placing volition&lt;/td&gt;
&lt;td&gt;Local sidecar, or the heuristic&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The quarterly letter&lt;/td&gt;
&lt;td&gt;Composed in the app. A model reply is kept only inside the word range, and only if it cites at least five known ids&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Ritual voice&lt;/td&gt;
&lt;td&gt;ElevenLabs if the key is set, otherwise a local tone and whisper.cpp&lt;/td&gt;
&lt;td&gt;The voice does, when that call is made. The Flutter screen still types the answer&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Web redaction&lt;/td&gt;
&lt;td&gt;The server&lt;/td&gt;
&lt;td&gt;Yes. The phone tier is the one that redacts before anything is sent&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;One local extract run on this hardware cost $0.00. I did not meter the same four sentences against a hosted chat API, so I will not invent a monthly bill. The number I have is the one in the table: $0, because no token was purchased.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why open weights matter here
&lt;/h2&gt;

&lt;p&gt;Purpose is the most personal corpus a person owns. A twelve-month model of you cannot live on someone else's server, behind someone else's model, on someone else's subscription. You also have to inspect why the app concluded something about your life. Open weights make both possible. Every text step in this build runs on those weights. There is no hosted LLM in that path. ElevenLabs is the one hosted model, and it covers voice only. Backboard is a comparison gateway, not the production extractor.&lt;/p&gt;

&lt;p&gt;The four circles need a credit, because they are not traditional ikigai. Andrés Zuzunaga drew a purpose diagram in 2011. Marc Winn relabelled it ikigai in 2014, and Dan Buettner's reporting from Okinawa carried that picture further. Traditional ikigai is smaller and more daily. purpoise keeps the diagram as one mode and a weekly practice as the other. Money is not required. The paid circle may stay empty, and the app is allowed to say the picture is unclear.&lt;/p&gt;

&lt;p&gt;The limits, from the privacy note, without the marketing gloss:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Web has no on-device redaction. Redacted text reaches the server. Mobile does not.&lt;/li&gt;
&lt;li&gt;The stores and the API host were designed as managed accounts. In this build the stores were not written, and the API host was not created.&lt;/li&gt;
&lt;li&gt;A tunnel, when one is used, transits Cloudflare. They see redacted text in flight.&lt;/li&gt;
&lt;li&gt;Production inference, when the API is up, depends on one laptop being awake.&lt;/li&gt;
&lt;li&gt;A search API would learn what you are curious about, at query-string granularity. That call was not made.&lt;/li&gt;
&lt;li&gt;ElevenLabs hears your voice for a ritual answer, when the route is called.&lt;/li&gt;
&lt;li&gt;Backboard saw an authenticated request and refused the chat. It did not receive a production corpus.&lt;/li&gt;
&lt;li&gt;Redaction reduces exposure. It does not eliminate it. Context can still imply a third party.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What I can count
&lt;/h2&gt;

&lt;p&gt;I did not run the app on my own week of mail and spending, so I will not report hours saved or decisions made. The measured use is the fixture set.&lt;/p&gt;

&lt;p&gt;The bake-off took four sentences and four local models. Selection was latency, not quality, because quality was tied. The held-out pair is two rows. The sidecar smoke test was one feature vector. On the local client the four circles meet, and the verdict is that the work carries it. The verdict page reads each question and the pairs that both have a trace. Rejecting the teaching claim opens the middle again, drops Mission and Vocation, and the verdict becomes partial. The recording above is that sequence. The engine fixture the tests lock is still the older partial case, with pay missing. That recompute is the lazy payoff I can stand behind: you do not re-score your life by hand. You reject one receipt and the conclusion moves.&lt;/p&gt;

&lt;p&gt;Time to payoff on the public client is the time to open &lt;a href="https://jedau.github.io/purpoise/" rel="noopener noreferrer"&gt;https://jedau.github.io/purpoise/&lt;/a&gt; and read the verdict. There is no account required for that demo ledger.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sessions
&lt;/h2&gt;

&lt;p&gt;A DevRelay session was not exported. An Entire session link was not captured. The decision log is &lt;code&gt;docs/DECISIONS/&lt;/code&gt;. The unknowns log is &lt;code&gt;docs/UNKNOWNS.md&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize categories
&lt;/h2&gt;

&lt;p&gt;No partner line from the original list passed its checklist, so none is claimed. The entry is the overall challenge. Removed, and why:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Gemma. The on-device file was not downloaded, and the parity run was not measured.&lt;/li&gt;
&lt;li&gt;TabPFN. The sidecar answers a numeric vector locally. Accuracy against a naive baseline was not reported for that task.&lt;/li&gt;
&lt;li&gt;Tinker. The export and a local delta ran. The fine-tune API was not called.&lt;/li&gt;
&lt;li&gt;Render. The blueprint is in the repo. Creating a service asked for a card even on the free web plan. The API was left undeployed.&lt;/li&gt;
&lt;li&gt;Auth0. Sign-in exists. Token Vault was not called, and there is no step-up gate.&lt;/li&gt;
&lt;li&gt;Backboard. The key authenticates. Chat was refused. The stored scores are the local bake-off. The assistant does not use it for memory.&lt;/li&gt;
&lt;li&gt;ElevenLabs. The server route can speak and transcribe. The screen in the client still types.&lt;/li&gt;
&lt;li&gt;Browserbase. A download the person already has is read from a file. No browser session is opened.&lt;/li&gt;
&lt;li&gt;Entire, GitHub Copilot, Mastra, MongoDB Atlas, Sentry, SerpApi, Temporal, Tiger Data. The code names the role. The checklist for each one asks for a run this build did not do.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;The repo is &lt;a href="https://github.com/jedau/purpoise" rel="noopener noreferrer"&gt;https://github.com/jedau/purpoise&lt;/a&gt;. The licence is MIT.&lt;/p&gt;

&lt;p&gt;Locally: copy &lt;code&gt;.env.example&lt;/code&gt; to &lt;code&gt;.env&lt;/code&gt;, then follow &lt;code&gt;docs/RUNBOOK.md&lt;/code&gt;. The web build publishes from &lt;code&gt;.github/workflows/deploy-pages.yml&lt;/code&gt; on a push to &lt;code&gt;main&lt;/code&gt;. The base path is &lt;code&gt;/purpoise/&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Commits after 2026-10-25 23:59 PDT, if any, are listed in the README.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Woolly: Gemma proposes the tasks. Only the client can pay for them.</title>
      <dc:creator>Jed Aureus Gonzales</dc:creator>
      <pubDate>Sun, 18 Oct 2026 03:02:44 +0000</pubDate>
      <link>https://dev.to/jedau/woolly-gemma-proposes-the-tasks-only-the-client-can-pay-for-them-1ip3</link>
      <guid>https://dev.to/jedau/woolly-gemma-proposes-the-tasks-only-the-client-can-pay-for-them-1ip3</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;Woolly is a board for a client who pays a few freelancers against a project budget.&lt;/p&gt;

&lt;p&gt;The client sets the budget, then puts an allocation on each cycle. Tasks on that project split the allocated money, and the shares always add up to 100%. A worker can edit a task and submit proof. Marking it complete requires a written summary of the proof, and that moves the task to &lt;strong&gt;Under Review&lt;/strong&gt;. The client is the only person who can approve it (&lt;strong&gt;Complete&lt;/strong&gt;) or send it back (&lt;strong&gt;Revision Needed&lt;/strong&gt;). A worker who tries to approve their own work is refused by the server.&lt;/p&gt;

&lt;p&gt;That is the thing someone would pay for. A client who already hires people by the piece wants the money to stay a percentage of a number they typed, with one approver. Woolly is not a marketplace and it does not send invoices. It is the agreement: budget, cycles, shares, proof, and a client who decides.&lt;/p&gt;

&lt;p&gt;In the walkthrough, Field notes has an $800 allocation and four tasks. Each one sits at 25%, $200. Move one slider and the others give up the remainder so the total stays 100%.&lt;/p&gt;

&lt;p&gt;Gemma does not touch that number. A client can paste a summary ("add the landing page to Cycle 1") and a Mastra agent asks Gemma to propose the tasks, on the cycle named in the summary. If that cycle is missing, the task comes back unplaced. The agent does not quietly pick a different cycle. A second agent can recommend credit or dispute on a proof. The recommendation is stored. The payment decision is a separate action, and only the client can take it.&lt;/p&gt;

&lt;p&gt;The people in the demo are fictional: Cool Client, Worker One, and Worker Dos. There is no password.&lt;/p&gt;

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

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/IX1FJzK4vcU" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;The video is the client board, a task and its allocation, Manage (budgets and cycles), a new project, then the same app as Worker One. The worker has no project button and no Settings. The account menu is Sign out. Nothing in the recording is saved.&lt;/p&gt;

&lt;p&gt;To run it locally, from a clone of the repo:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;cd &lt;/span&gt;server &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; npm &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; npm run migrate &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; npm run dev
&lt;span class="nb"&gt;cd &lt;/span&gt;app &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; flutter run &lt;span class="nt"&gt;-d&lt;/span&gt; chrome
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Copy &lt;code&gt;.env.example&lt;/code&gt; to &lt;code&gt;.env&lt;/code&gt; first. The API listens on &lt;code&gt;http://127.0.0.1:8787&lt;/code&gt;, and the app is the Flutter web client. &lt;code&gt;npm run migrate&lt;/code&gt; rebuilds the tables and the fictional rows. Sign in as &lt;strong&gt;Cool Client&lt;/strong&gt; or &lt;strong&gt;Worker One&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The board loads without a model call. Gemma runs when someone asks for tasks from a summary, or for a recommendation on a proof.&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/jedau" rel="noopener noreferrer"&gt;
        jedau
      &lt;/a&gt; / &lt;a href="https://github.com/jedau/woolly" rel="noopener noreferrer"&gt;
        woolly
      &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;Woolly&lt;/h1&gt;
&lt;/div&gt;

&lt;p&gt;Pay for an iteration, measured by the subtasks a client accepts.&lt;/p&gt;

&lt;p&gt;A project is the epic. An iteration is one sprint inside that project: it has an amount and an optional deadline. Tasks are the stories. Subtasks are the tickets. Pay for an iteration is its amount times the share of subtask effort the client has credited.&lt;/p&gt;

&lt;p&gt;A client has many workers. A worker has many projects. The client adds work and reviews it, filtered by worker or by project. The worker adds and edits tasks, and submits subtasks for review. The worker cannot credit their own work.&lt;/p&gt;

&lt;p&gt;The people in the demo are fictional: Cool Client (client), Worker One (worker), and Worker Dos (worker).&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Demo&lt;/h2&gt;
&lt;/div&gt;

&lt;p&gt;Walkthrough of the client board, task allocation, Manage, and the worker account: &lt;a href="https://youtu.be/IX1FJzK4vcU" rel="nofollow noopener noreferrer"&gt;https://youtu.be/IX1FJzK4vcU&lt;/a&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Look&lt;/h2&gt;
&lt;/div&gt;

&lt;p&gt;Dark brown, flat, closer to a Darcula editor than a paper form. Near-black background, warm brown panels, amber marks…&lt;/p&gt;&lt;/div&gt;


&lt;/div&gt;
&lt;br&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/jedau/woolly" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;br&gt;
&lt;/div&gt;
&lt;br&gt;


&lt;p&gt;MIT. &lt;a href="https://github.com/jedau/woolly" rel="noopener noreferrer"&gt;github.com/jedau/woolly&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;The open pieces are Gemma and Mastra. Gemma is &lt;code&gt;gemma-4-26b-a4b-it&lt;/code&gt;, called through the Gemini API (&lt;code&gt;server/src/env.ts&lt;/code&gt;). It is an open-weight model on a hosted endpoint. This build does not run the weights on the laptop, and I am not going to pretend it does. Proof text and summaries go to that API when an agent runs.&lt;/p&gt;

&lt;p&gt;Mastra is the harness (&lt;code&gt;server/src/woolly.ts&lt;/code&gt;):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Agent&lt;/th&gt;
&lt;th&gt;Job&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Place&lt;/td&gt;
&lt;td&gt;Turn a summary into one task per piece of work, on the cycle the summary names&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Split&lt;/td&gt;
&lt;td&gt;Break a brief into tasks and subtasks. Effort is the size, value is how much it matters&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Review&lt;/td&gt;
&lt;td&gt;Read a proof against the description and recommend credit or dispute&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Place and split return JSON the server checks before anything is written. Review writes &lt;code&gt;model_outcome&lt;/code&gt; on the proof. The client's decision is a different request. Shares are basis points on the task row, and the server normalizes them so a project sums to 10000. Cycle allocations that would pass the project budget are rejected. None of that is a sentence in the prompt.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;search-past-work&lt;/code&gt; is a Mastra tool the split and review agents call. It looks up this project's own accepted subtasks, with Postgres full text and a pgvector embedding, and hands a few titles back. The embedding call is Gemini's embedding model, not Gemma. The memory is the client's past work, not a general chat log.&lt;br&gt;
&lt;/p&gt;

&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart LR
  summary[Client summary] --&amp;gt; place[Mastra place agent]
  place --&amp;gt; gemma[Gemma via the Gemini API]
  gemma --&amp;gt; tasks[Tasks on the named cycle]
  tasks --&amp;gt; board[Board and allocation]
  proof[Worker proof] --&amp;gt; review[Mastra review agent]
  review --&amp;gt; note[Recommendation only]
  note --&amp;gt; client[Client approves or sends it back]
  client --&amp;gt; paid[Share of the allocation]&lt;/code&gt;&lt;/pre&gt;



&lt;p&gt;Stack around that: a Flutter web client, a TypeScript API, Postgres. The theme is a dark brown board. Projects, workers, cycle, and status are filters. The client and the worker are two workplaces in one app.&lt;/p&gt;

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

&lt;p&gt;A client will pay for this only if the money rule is boring and visible. An allocation is a number they set. A share is a percentage of that number. A worker cannot approve the proof. If those rules lived inside a closed chat product, the same vendor would propose the work and hold the ledger, and a prompt would be the only thing standing between a suggestion and a payment.&lt;/p&gt;

&lt;p&gt;Other people building agents next to money are drawing the same line. A spending limit written in a prompt is not a control (&lt;a href="https://dev.to/payload-tools/prompts-dont-enforce-budgets-signatures-do-2a1j"&gt;Prompts don't enforce budgets. Signatures do.&lt;/a&gt;), and the approval has to sit outside the model (&lt;a href="https://dev.to/quinn_854b15f517d8632ed4f/should-this-ai-agent-be-allowed-to-pay-designing-the-approval-layer-between-agents-and-company-20k7"&gt;Should this AI agent be allowed to pay?&lt;/a&gt;). Woolly keeps Gemma on the proposal. The budget check, the shares, and the approver check are ordinary server code.&lt;/p&gt;

&lt;p&gt;Open is what makes the split possible. Mastra is the harness, so the three agents are a file I can read, and the model is a setting (&lt;code&gt;GEMINI_MODEL&lt;/code&gt;, default &lt;code&gt;gemma-4-26b-a4b-it&lt;/code&gt;) rather than the product. I can point Place at another Gemma endpoint without rewriting who is allowed to mark a task paid. The board also has a price of zero when nobody is asking the model anything. The calls a client would pay for are the ones that propose work or read a proof.&lt;/p&gt;

&lt;p&gt;What open does not do here: it does not keep the proof on the machine. Hosted Gemma means a summary leaves the laptop when an agent runs. Local weights would close that gap. They are not what this submission runs.&lt;/p&gt;

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

&lt;p&gt;Featured&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma.&lt;/strong&gt; Place, split, and review are Gemma (&lt;code&gt;gemma-4-26b-a4b-it&lt;/code&gt; through the Gemini API). Place turns a summary into tasks. Review recommends credit or dispute and does not get to apply it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Partner&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Mastra.&lt;/strong&gt; Those three agents are Mastra &lt;code&gt;Agent&lt;/code&gt;s, and past work is a &lt;code&gt;createTool&lt;/code&gt; the split and review agents call before they answer.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Solo entry.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
    </item>
    <item>
      <title>LaunchPup: rest is the countdown, the walk is liftoff (for both of you)</title>
      <dc:creator>Jed Aureus Gonzales</dc:creator>
      <pubDate>Sun, 11 Oct 2026 07:02:32 +0000</pubDate>
      <link>https://dev.to/jedau/launchpup-rest-is-the-countdown-the-walk-is-liftoff-for-both-of-you-3450</link>
      <guid>https://dev.to/jedau/launchpup-rest-is-the-countdown-the-walk-is-liftoff-for-both-of-you-3450</guid>
      <description>&lt;p&gt;Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass Submission 🌿&lt;/p&gt;

&lt;p&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;/p&gt;

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

&lt;p&gt;LaunchPup is mission control for a walk you take &lt;em&gt;with&lt;/em&gt; your dog.&lt;/p&gt;

&lt;p&gt;Most “walk the dog” apps treat the human as a dispatcher: log the pee, close the ticket, get back to the couch. That is not a walk. When the leash clips on, two mammals leave the house. The dog is the excuse. The owner is the payload. Touch Grass this week is about getting a person off the screen; LaunchPup uses the nap log as the countdown so that person actually goes.&lt;/p&gt;

&lt;p&gt;You log rest and the last burn. &lt;a href="https://github.com/PriorLabs/tabpfn" rel="noopener noreferrer"&gt;TabPFN&lt;/a&gt; reads a small CSV of rest/walk features and votes &lt;strong&gt;HOLD / GO / OVERDUE&lt;/strong&gt;. Range safety votes separately on pavement temperature, humidity, night, and short-nosed breeds. &lt;a href="https://ai.google.dev/gemma" rel="noopener noreferrer"&gt;Gemma&lt;/a&gt; is CAPCOM only: it does not get to overturn the poll. It writes a sixty-second pocket card, then tells you to put the phone away. &lt;a href="https://mastra.ai" rel="noopener noreferrer"&gt;Mastra&lt;/a&gt; is the harness around that agent.&lt;/p&gt;

&lt;p&gt;Then the UI becomes a countdown with one button left: splashdown. That clock is for both of you on the sidewalk.&lt;br&gt;
&lt;/p&gt;

&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart LR
  naps[Nap / walk log] --&amp;gt; tabpfn[TabPFN energy vote]
  weather[Open-Meteo + pavement] --&amp;gt; range[Range safety]
  tabpfn --&amp;gt; ld[Launch director]
  range --&amp;gt; ld
  ld --&amp;gt; gemma[Gemma CAPCOM via Mastra]
  gemma --&amp;gt; pocket[Owner pockets the phone and walks]&lt;/code&gt;&lt;/pre&gt;



&lt;p&gt;The open pieces are the product, not garnish. TabPFN never sees a prompt. It sees twelve numbers (rest in 24h, nap count, longest nap, hours since the last walk, last walk length, age, weight, breed energy, hour, air temp, humidity, brachy flag). Gemma is forbidden from changing HOLD to GO because the pavement is 52°C. Range safety is boring code on purpose: a foundation model should not be the thing that decides whether asphalt will fry a pad.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Code:&lt;/strong&gt; &lt;a href="https://github.com/jedau/launchpup" rel="noopener noreferrer"&gt;github.com/jedau/launchpup&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Live:&lt;/strong&gt; run it locally (below). No login, no account. A short screencast of poll → T-0 → Liftoff counts as the required demo if you do not deploy.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;cp&lt;/span&gt; .env.example .env
uv &lt;span class="nb"&gt;sync
&lt;/span&gt;uv run python scripts/generate_training.py
uv run uvicorn launchpup.app:app &lt;span class="nt"&gt;--reload&lt;/span&gt; &lt;span class="nt"&gt;--port&lt;/span&gt; 8000
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Open &lt;a href="http://127.0.0.1:8000" rel="noopener noreferrer"&gt;http://127.0.0.1:8000&lt;/a&gt;. Scout (the default dog) already looks overdue, so the first &lt;strong&gt;Poll the room&lt;/strong&gt; is a real count, not an empty form. &lt;strong&gt;T-0 call&lt;/strong&gt; prints the pocket card. &lt;strong&gt;Liftoff&lt;/strong&gt; starts the walk clock and stops offering more screen. &lt;strong&gt;Splashdown&lt;/strong&gt; is you both coming home.&lt;/p&gt;

&lt;p&gt;Optional: &lt;code&gt;TABPFN_TOKEN&lt;/code&gt; for the real tabular vote, Ollama &lt;code&gt;gemma2:2b&lt;/code&gt; plus &lt;code&gt;npm run capcom&lt;/code&gt; for Gemma through Mastra, or &lt;code&gt;GROQ_API_KEY&lt;/code&gt; if you have no local GPU.&lt;/p&gt;

&lt;h3&gt;
  
  
  How it keeps the screen the shortest part
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;The decision is a poll, not a chat. You do not negotiate with a chatbot about whether to go outside.&lt;/li&gt;
&lt;li&gt;CAPCOM is sixty seconds of card, then a hard pocket mode.&lt;/li&gt;
&lt;li&gt;Liftoff has nothing left to tap except splashdown.&lt;/li&gt;
&lt;li&gt;The owner is addressed as the walker. The dog does not “get walked” by an app.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why open innovation matters
&lt;/h2&gt;

&lt;p&gt;The prompt asked where open is better than closed. For a walk app, closed wants your nap times on someone else's server and your face in the feed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The fueling log stays on the machine.&lt;/strong&gt; Rest and walks live in local SQLite. There is no account. A closed walk tracker is a surveillance product with a cartoon bone on it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The energy vote is a table you can read.&lt;/strong&gt; TabPFN (Prior Labs) classifies from &lt;code&gt;data/zoomie_windows.csv&lt;/code&gt;. Training labels are synthetic and generated in-repo — I am not pretending this is a vet study. You can inspect the rows, disagree with HOLD, and still go. If TabPFN's local license token is missing, the UI says so and falls back to a labeled sklearn classifier instead of faking “TabPFN.” Honesty is part of the open bet.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;CAPCOM is a model you can swap.&lt;/strong&gt; Gemma via Ollama on a laptop, Groq's &lt;code&gt;gemma2-9b-it&lt;/code&gt; if you have no GPU, or a printed fallback card if the network is gone. Mastra orchestrates the agent; it does not own the weights. I can change how CAPCOM talks without retraining the thing that decides GO.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It can cost nothing to run.&lt;/strong&gt; Local TabPFN + local Gemma + Open-Meteo for weather. Open-Meteo is the one network call range safety wants, because pavement physics is not a personality.&lt;/p&gt;

&lt;p&gt;Where open worked better than closed: I could put the &lt;em&gt;decision&lt;/em&gt; in a foundation model built for tables, keep the &lt;em&gt;veto&lt;/em&gt; in code, and keep the &lt;em&gt;narration&lt;/em&gt; in an open-weight LLM — three stations, votes in the clear, phone in a pocket. A single closed chat API would have mixed those jobs and then asked me to keep scrolling.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it's put together
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Station&lt;/th&gt;
&lt;th&gt;Job&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;TabPFN&lt;/td&gt;
&lt;td&gt;Energy: HOLD / GO / OVERDUE from rest history&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Range safety&lt;/td&gt;
&lt;td&gt;Pavement, heat, humidity, night, brachy — can scrub a GO&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Launch director&lt;/td&gt;
&lt;td&gt;Combine the two. Overdue + hot asphalt is still a scrub&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemma via Mastra&lt;/td&gt;
&lt;td&gt;Pocket card only. Does not change the votes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pocket clock&lt;/td&gt;
&lt;td&gt;The actual product&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Stack: FastAPI + a small mission-control page, SQLite, Python 3.12, a Mastra agent/workflow on Node for CAPCOM.&lt;/p&gt;

&lt;p&gt;Credits that are not mine: TabPFN, Gemma, Mastra, Open-Meteo, FastAPI, scikit-learn (fallback).&lt;/p&gt;

&lt;p&gt;Not veterinary advice. If the pavement fries an egg, it will fry a paw — and it is a bad hour for the human on the other end of the leash.&lt;/p&gt;

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

&lt;p&gt;Featured&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of TabPFN&lt;/strong&gt; — the energy vote is a &lt;code&gt;TabPFNClassifier&lt;/code&gt; on a rest/walk CSV (&lt;code&gt;launchpup/tabpfn_engine.py&lt;/code&gt;). This is the launch window, not a sidebar chart.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt; — CAPCOM writes the T-0 pocket card from the poll (Ollama or Groq &lt;code&gt;gemma2*&lt;/code&gt;). Gemma narrates; it does not get a veto.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Partner&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Mastra&lt;/strong&gt; — &lt;code&gt;agent/&lt;/code&gt; CAPCOM &lt;code&gt;Agent&lt;/code&gt; plus a &lt;code&gt;t0-brief&lt;/code&gt; workflow that turns the poll into the card.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Momentum: An AI-Powered Activity Prioritizer for Analysis Paralysis</title>
      <dc:creator>Jed Aureus Gonzales</dc:creator>
      <pubDate>Mon, 05 Oct 2026 06:50:39 +0000</pubDate>
      <link>https://dev.to/jedau/momentum-an-ai-powered-activity-prioritizer-for-analysis-paralysis-2ogh</link>
      <guid>https://dev.to/jedau/momentum-an-ai-powered-activity-prioritizer-for-analysis-paralysis-2ogh</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-weekend-2026-10-01"&gt;Hacktoberfest Weekend Challenge: Build for a Friend&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;I built &lt;strong&gt;Momentum&lt;/strong&gt; for my wife, who struggles with analysis paralysis when facing multiple competing priorities. She describes it as "having 47 browser tabs open in my brain" — every task feels equally urgent, every deadline feels immediate, and the mental load of choosing becomes so overwhelming that nothing gets done.&lt;/p&gt;

&lt;p&gt;Momentum is a voice-first, AI-powered activity prioritizer that helps her break through decision paralysis by:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Capturing tasks via voice&lt;/strong&gt; (no typing friction when she's already overwhelmed)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Applying multi-factor heuristics&lt;/strong&gt; to score each task across dimensions like urgency, importance, energy required, time available, and emotional weight&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Explaining its reasoning&lt;/strong&gt; in plain language so she can trust (and override) the recommendation&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Suggesting the single smallest next action&lt;/strong&gt; for the top task, not the whole project&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Learning from her feedback&lt;/strong&gt; to improve future recommendations&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The app doesn't just dump a sorted list back on her — that would be another source of overwhelm. Instead, it presents one clear recommendation with a "why," gives her agency to accept, defer, or reshuffle, and celebrates the small wins to build momentum.&lt;/p&gt;

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

&lt;p&gt;The demo shows:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Voice input via ElevenLabs for task capture&lt;/li&gt;
&lt;li&gt;Real-time prioritization using Gemma 2B running locally&lt;/li&gt;
&lt;li&gt;The "One Thing" view that hides everything except the current recommendation&lt;/li&gt;
&lt;li&gt;Energy-aware scheduling (won't suggest high-energy tasks when she marks herself as drained)&lt;/li&gt;
&lt;li&gt;Feedback loop where she can rate recommendations to improve future scoring&lt;/li&gt;
&lt;/ul&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/jedau" rel="noopener noreferrer"&gt;
        jedau
      &lt;/a&gt; / &lt;a href="https://github.com/jedau/momentum" rel="noopener noreferrer"&gt;
        momentum
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
       A voice-first, AI-assisted activity prioritizer for people who get stuck when everything feels equally urgent.
    &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;Momentum&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;A voice-first, AI-assisted activity prioritizer for people who get stuck when
everything feels equally urgent. Instead of returning a sorted list (which is
just more overwhelm), Momentum picks &lt;strong&gt;one&lt;/strong&gt; task, explains why, and suggests
the smallest next action.&lt;/p&gt;
&lt;p&gt;Built for the &lt;a href="https://dev.to/challenges/hacktoberfest-weekend-2026-10-01" rel="nofollow"&gt;Hacktoberfest Weekend Challenge: Build for a Friend&lt;/a&gt;.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;What it does&lt;/h2&gt;
&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Capture by voice or text&lt;/strong&gt; — no form required. A parser infers urgency
importance, energy demand, duration, and emotional weight from plain language.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multi-factor scoring&lt;/strong&gt; — urgency, importance, energy match, time fit, and
learned history, combined with transparent weights.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One Thing view&lt;/strong&gt; — hides everything except the current recommendation, with
a plain-language explanation you can inspect (per-factor breakdown included).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Smallest next action&lt;/strong&gt; — turns "organize the garage" into "pick up one thing
and decide keep / trash / donate".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Agency and feedback&lt;/strong&gt; — accept, defer, reshuffle, mark done, or flag "too
big". Those choices feed the history…&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/jedau/momentum" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The repository includes:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Full Next.js application with TypeScript&lt;/li&gt;
&lt;li&gt;Docker setup for local Gemma inference via Ollama&lt;/li&gt;
&lt;li&gt;Mastra agent configuration for orchestration&lt;/li&gt;
&lt;li&gt;MongoDB Atlas schema and vector search indexes&lt;/li&gt;
&lt;li&gt;Temporal workflow definitions for durable task processing&lt;/li&gt;
&lt;li&gt;Sentry configuration for agent tracing&lt;/li&gt;
&lt;/ul&gt;

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

&lt;h3&gt;
  
  
  Architecture Overview
&lt;/h3&gt;

&lt;p&gt;Momentum is built around open-source AI at every layer:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Local Inference with Gemma (Google)&lt;/strong&gt;&lt;br&gt;
I run Google's open-weight Gemma 2B model locally via Ollama for all NLP tasks — parsing voice transcripts into structured tasks, generating the "smallest next action" suggestions, and explaining prioritization reasoning. Running locally means my wife's task list never leaves our network, which matters for personal data like medical appointments or sensitive work tasks.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Agent Orchestration with Mastra&lt;/strong&gt;&lt;br&gt;
The core prioritization logic is wrapped in a Mastra agent workflow that:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Ingests tasks from voice or text input&lt;/li&gt;
&lt;li&gt;Applies scoring heuristics (Eisenhower Matrix + Energy Accounting)&lt;/li&gt;
&lt;li&gt;Generates explanations using Gemma&lt;/li&gt;
&lt;li&gt;Handles user feedback loops&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Mastra's tool-calling interface let me connect the agent to MongoDB for persistence and to Temporal for durable execution.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Vector Memory with MongoDB Atlas&lt;/strong&gt;&lt;br&gt;
I use MongoDB Atlas Vector Search to store task embeddings and historical patterns. When my wife adds a new task, the system searches for similar past tasks to learn what she actually prioritized in comparable situations — not just what she said was important. This catches patterns like "I always defer the 'important but not urgent' health tasks even when I mark them high priority."&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Voice Interface with ElevenLabs&lt;/strong&gt;&lt;br&gt;
The voice capture uses ElevenLabs for transcription, but the key feature is the &lt;strong&gt;voice persona&lt;/strong&gt; — I fine-tuned a calm, non-judgmental voice that reads recommendations back to her. When you're already feeling overwhelmed by your to-do list, a harsh robotic voice makes it worse. This one sounds like a patient friend.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Durable Workflows with Temporal&lt;/strong&gt;&lt;br&gt;
Task prioritization isn't instant — it involves embedding generation, model inference, and database writes. I wrapped the entire flow in a Temporal workflow so if her laptop dies mid-processing, the workflow resumes exactly where it left off. No lost tasks, no duplicate entries.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Deployment on Render&lt;/strong&gt;&lt;br&gt;
The app is deployed on Render with a Web Service for the Next.js frontend and a Background Worker for the Temporal workflows. Render's native support for long-running processes made it perfect for the agent orchestration layer.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Agent Tracing with Sentry&lt;/strong&gt;&lt;br&gt;
I instrumented the Mastra agent with Sentry's agent tracing to see exactly how many tokens each prioritization request uses, where latency spikes happen, and how often the model generates invalid outputs. This helped me tune the Gemma prompts and catch when the local model was struggling with complex multi-task inputs.&lt;/p&gt;
&lt;h3&gt;
  
  
  The Prioritization Algorithm
&lt;/h3&gt;

&lt;p&gt;The scoring combines explicit user inputs with learned patterns:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Priority Score = (
0.30 × Eisenhower_Urgency +
0.25 × Eisenhower_Importance +
0.20 × Energy_Match (task energy vs. current energy) +
0.15 × Time_Fit (task duration vs. available window) +
0.10 × Historical_Pattern (what she actually did with similar tasks)
)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Gemma generates the "smallest next action" by taking the top-scored task and prompting:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Given this task and her current energy level, what is the absolute smallest first step she could take in under 5 minutes? Be specific and encouraging."&lt;/p&gt;
&lt;/blockquote&gt;

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

&lt;p&gt;&lt;strong&gt;Privacy by Design&lt;/strong&gt;&lt;br&gt;
My wife's task list includes sensitive things — doctor appointments, work conflicts, family concerns. Running Gemma locally via Ollama means that data never hits a third-party API. The voice transcripts are processed through ElevenLabs (which she opted into), but the actual task content, priorities, and patterns stay on our hardware or in our MongoDB Atlas cluster.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Inspectability and Trust&lt;/strong&gt;&lt;br&gt;
Analysis paralysis thrives on opacity. If an app says "do this task" with no explanation, she'll second-guess it into oblivion. Because Gemma is open-weight, I can inspect exactly how it generates explanations and tune the prompts to be more transparent. She can ask "why this task?" and get a real answer grounded in the scoring heuristics.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Adaptability Without Vendor Lock-in&lt;/strong&gt;&lt;br&gt;
If Gemma 2B starts feeling too small for her growing task history, I can swap in Qwen 3, Llama 3.2, or any other open-weight model without rewriting the application. The Mastra orchestration layer is model-agnostic. Compare that to being locked into GPT-4's API pricing and rate limits — for a personal tool that might run for years, that matters.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Cost Sustainability&lt;/strong&gt;&lt;br&gt;
A hosted API would cost $0.01-0.03 per prioritization request. That sounds small until she's capturing 20+ tasks daily and iterating on priorities throughout the day. Running Gemma locally on her existing laptop costs nothing per request, making the tool sustainable for daily use without budget anxiety.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Community Improvement&lt;/strong&gt;&lt;br&gt;
Because every component is open-source, I can contribute back. I filed a small PR with Mastra to improve their Ollama integration, and I'm documenting the Gemma fine-tuning process for voice-tone adaptation. The tool improves as the ecosystem improves.&lt;/p&gt;

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

&lt;p&gt;I'm entering the following partner categories:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt; — Core prioritization and explanation generation runs on Google's open-weight Gemma 2B model&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Mastra&lt;/strong&gt; — Agent orchestration, tool calling, and workflow management built on Mastra&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of MongoDB Atlas&lt;/strong&gt; — Vector search for task similarity, long-term memory for patterns, and document storage&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of ElevenLabs&lt;/strong&gt; — Voice transcription and text-to-speech for the voice-first interface&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Temporal&lt;/strong&gt; — Durable workflows for task processing that survive interruptions&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render&lt;/strong&gt; — Web service and background worker deployment&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Sentry Agent Tracing&lt;/strong&gt; — Instrumented agent performance monitoring and token/latency tracking&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Appendix: What She Said
&lt;/h2&gt;

&lt;p&gt;I handed Momentum to my wife on Sunday morning. By Sunday evening, she said:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"I don't know how to explain it, but having it pick ONE thing and tell me why in a nice voice... I actually did three tasks today. Usually I'd have done none just thinking about all of them."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The "smallest next action" feature was the breakthrough. Instead of "organize the garage" (overwhelming), it suggested "pick up one thing from the garage floor and decide keep/trash/donate." That she could do. And once she did one, she did another. That's the momentum.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>weekendchallenge</category>
      <category>hf26challenge</category>
    </item>
  </channel>
</rss>
