<?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: HarshadaGirase</title>
    <description>The latest articles on DEV Community by HarshadaGirase (@harshadagirase).</description>
    <link>https://dev.to/harshadagirase</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%2F1367913%2F54194f34-a6d6-4999-a926-fa92bd7d3787.jpeg</url>
      <title>DEV Community: HarshadaGirase</title>
      <link>https://dev.to/harshadagirase</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/harshadagirase"/>
    <language>en</language>
    <item>
      <title>Ghosts in node_modules: I trained an open model to predict which packages will be abandoned</title>
      <dc:creator>HarshadaGirase</dc:creator>
      <pubDate>Sun, 01 Nov 2026 07:02:06 +0000</pubDate>
      <link>https://dev.to/harshadagirase/ghosts-in-nodemodules-i-trained-an-open-model-to-predict-which-packages-will-be-abandoned-8l4</link>
      <guid>https://dev.to/harshadagirase/ghosts-in-nodemodules-i-trained-an-open-model-to-predict-which-packages-will-be-abandoned-8l4</guid>
      <description>&lt;h1&gt;
  
  
  DepForecast: A Weather Forecast for the Ghosts in Your &lt;code&gt;node_modules&lt;/code&gt; (work in progress)
&lt;/h1&gt;

&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;Abandoned dependencies are ghosts in the machine: code whose maintainers are gone but which still runs in production. Developers usually find out only after something breaks or a vulnerability goes unpatched.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;DepForecast&lt;/strong&gt; is a weather forecast for your dependencies. It's a local, open-source CLI that will predict which of a project's npm dependencies are likely to be abandoned in the next 12 months, explain why, and suggest ways to help keep them alive.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Renovate keeps you current. DepForecast tells you when "current" is about to stop existing.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The theme shows up in the output design: 🪦 &lt;strong&gt;RESTING&lt;/strong&gt; for dead packages (with a short, factual obituary), 👻 &lt;strong&gt;HAUNTING&lt;/strong&gt; for at-risk ones, and 🌼 &lt;strong&gt;ALIVE&lt;/strong&gt; for healthy ones. There's also an &lt;em&gt;ofrenda&lt;/em&gt; section with ways to help: funding links, "help wanted" issues and contributing guides.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Status: work in progress.&lt;/strong&gt; I built the data foundation first, because a prediction is only as honest as the data behind it. What's done and tested:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Collectors for the npm registry and the GitHub GraphQL API, with a disk cache, retries and rate-limit handling&lt;/li&gt;
&lt;li&gt;A &lt;strong&gt;leakage guard&lt;/strong&gt; that blocks any training query from looking at data after its snapshot date&lt;/li&gt;
&lt;li&gt;A real dataset universe: &lt;strong&gt;250 popular JS/TS repos → 5,118 distinct dependencies → 3,719 GitHub repos&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The TabPFN predictor, the backtest, the &lt;code&gt;scan&lt;/code&gt; command and the Gemma explainer are designed in &lt;code&gt;PLAN.md&lt;/code&gt; and are next.&lt;/p&gt;

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

&lt;p&gt;No deployed link yet, since it's a local CLI. Here's what runs today:&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%2Fsxbev2t24bedsfflu6h2.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%2Fsxbev2t24bedsfflu6h2.png" alt=" " width="798" height="87"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2e7frnenp8zqb90ic8mi.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%2F2e7frnenp8zqb90ic8mi.png" alt=" " width="798" height="77"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The real dataset&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%2Fgwlou5pdulxwie4na97f.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%2Fgwlou5pdulxwie4na97f.png" alt=" " width="800" height="600"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The leakage guard working&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%2Fu6azqy397gkr8inouy2c.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%2Fu6azqy397gkr8inouy2c.png" alt=" " width="800" height="478"&gt;&lt;/a&gt;&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/HarshadaGirase" rel="noopener noreferrer"&gt;
        HarshadaGirase
      &lt;/a&gt; / &lt;a href="https://github.com/HarshadaGirase/DepForecast" rel="noopener noreferrer"&gt;
        DepForecast
      &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;DepForecast&lt;/h1&gt;

&lt;/div&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/HarshadaGirase/DepForecast" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;br&gt;
&lt;/div&gt;
&lt;br&gt;


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

&lt;p&gt;&lt;strong&gt;Open-source AI (planned core):&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;TabPFN-3.5 (open weights, run locally)&lt;/strong&gt; is the predictor. It's a tabular foundation model that learns in-context from &lt;code&gt;data/train.csv&lt;/code&gt;, which suits a small, honest dataset where training a deep model from scratch would overfit. There's a gradient-boosting fallback, so the tool never fails without the weights.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Gemma (&lt;code&gt;gemma3:4b&lt;/code&gt;) via Ollama, run locally&lt;/strong&gt;, will turn a risk score into a plain-English explanation. Its guardrails: every reason must cite a URL from the evidence (issue titles, README, release notes), and any suggested alternative must exist on npm with a release in the last 12 months. If Ollama is offline, it falls back to rule-based text.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;What's built (Python 3.11, Typer, Rich, Pydantic, httpx, pytest):&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Temporal validation, no leakage:&lt;/strong&gt; train on snapshots at 2024-10-01 and test on 2025-04-01, each with a 12-month outcome window. Even the seed repos' &lt;code&gt;package.json&lt;/code&gt; files are read &lt;em&gt;as they were&lt;/em&gt; on 2024-10-01, so the population itself doesn't leak the future.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The label, defined honestly:&lt;/strong&gt; abandoned = the repo was archived during the window, &lt;strong&gt;or&lt;/strong&gt; it had zero commits &lt;strong&gt;and&lt;/strong&gt; zero npm releases in the whole window. npm's &lt;code&gt;deprecated&lt;/code&gt; flag has no timestamp, so it's kept as a column, never as the label.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Features:&lt;/strong&gt; commits in the last 3/6/12 months, trend, distinct authors, bus factor (the top author's share of commits), days since last commit and release, issues opened vs. closed, stale PRs, repo age and funding. All are computed strictly from data before the snapshot date.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Efficient and free:&lt;/strong&gt; batched GraphQL costs about 1 point per batch of up to 10 repos, and every response is cached on disk, so reruns are free and work offline.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Monorepos:&lt;/strong&gt; collapsed to one row per repo (209 of them), and &lt;code&gt;@types/*&lt;/code&gt; is excluded.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I built it with an AI coding agent (Claude Code), working in small steps: 22 commits, each with tests.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Your dependency list is a map of your attack surface.&lt;/strong&gt; DepForecast runs locally and never sends it to a third-party service.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It's free to run on every PR&lt;/strong&gt;, so teams don't ration checks. The whole stack is free: the GitHub API, the npm registry, local TabPFN, local Gemma and GitHub Actions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A tool that warns about abandoned software shouldn't itself depend on a service that can disappear.&lt;/strong&gt; Open weights on your own machine can't be shut down or repriced.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Verifiable claims:&lt;/strong&gt; the features, the label definition and the data tables are in the repo. You can inspect them, retrain on your ecosystem and check every number.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Honest caveat:&lt;/strong&gt; TabPFN's weights are open but under a &lt;strong&gt;non-commercial&lt;/strong&gt; license, and Gemma follows Google's Gemma terms. Both are documented in &lt;code&gt;THIRD_PARTY.md&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;[Link or embed your DevRelay session here]&lt;/p&gt;

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

&lt;p&gt;While testing the collector, I asked GitHub about &lt;code&gt;request&lt;/code&gt;, one of the most famous packages in npm history. The answer was quiet: last commit &lt;strong&gt;February 11, 2020&lt;/strong&gt;. Zero commits in the year before my snapshot date, and zero after. It was never archived. The repo is still there and still looks like a living project, but npm now says &lt;em&gt;"request has been deprecated."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;That's the ghost: code that looks alive while nobody is home. DepForecast exists so the next &lt;code&gt;request&lt;/code&gt; shows up as 👻 HAUNTING in your terminal &lt;em&gt;before&lt;/em&gt; it becomes 🪦 RESTING in your production logs. And the ofrenda section is there so developers can do something about it while the maintainers are still around.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of TabPFN.&lt;/strong&gt; DepForecast is designed around &lt;strong&gt;TabPFN-3.5 (open weights, run locally)&lt;/strong&gt; as its predictor: it learns in-context from historical snapshots and outputs P(abandoned within 12 months). The pipeline that produces TabPFN's training set is built and tested: temporal windows, the label definition, the leakage guard and the 3,719-repo universe. &lt;code&gt;depforecast doctor&lt;/code&gt; already verifies the TabPFN license and token setup. &lt;em&gt;Status: the TabPFN predictor and its backtest against baselines are the next milestone.&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma.&lt;/strong&gt; &lt;strong&gt;&lt;code&gt;gemma3:4b&lt;/code&gt;, running locally via Ollama&lt;/strong&gt;, is the planned explainer. It will turn a risk score into evidence-cited advice, with a guardrail that drops any reason without a source URL. &lt;code&gt;depforecast doctor&lt;/code&gt; verifies that Ollama is running and &lt;code&gt;gemma3:4b&lt;/code&gt; is available. &lt;em&gt;Status: the explainer integration is planned.&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of GitHub Copilot (GitHub Actions).&lt;/strong&gt; &lt;strong&gt;GitHub Actions&lt;/strong&gt; automates the project: every push runs ruff, mypy, all 125 tests, and a guard that blocks committing model weights or files over 50 MB (&lt;code&gt;.github/workflows/ci.yml&lt;/code&gt;). The planned DepForecast Action will comment on PRs that add risky dependencies. &lt;em&gt;Note: I built this with Claude Code rather than Copilot, and I'm entering for the GitHub Actions automation.&lt;/em&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
    <item>
      <title>I stopped hunting for good first issues. A Gemma inside GitHub Actions does it while I sleep</title>
      <dc:creator>HarshadaGirase</dc:creator>
      <pubDate>Mon, 26 Oct 2026 07:20:01 +0000</pubDate>
      <link>https://dev.to/harshadagirase/i-stopped-hunting-for-good-first-i-stopped-hunting-for-good-first-issues-a-gemma-inside-github-3k8p</link>
      <guid>https://dev.to/harshadagirase/i-stopped-hunting-for-good-first-i-stopped-hunting-for-good-first-issues-a-gemma-inside-github-3k8p</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;p&gt;&lt;strong&gt;FirstPick&lt;/strong&gt; finds beginner-friendly GitHub issues in open-source AI repos, drops the ones someone has already taken, and hands me the 5 best, explained in plain English by Gemma running on my own laptop.&lt;/p&gt;

&lt;p&gt;I wanted my first contribution to an AI library. Searching "good first issue" meant ten tabs, and most of them were closed, assigned, already had a pull request, or had a "can I work on this?" comment from weeks ago. FirstPick does that digging for me:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Collects&lt;/strong&gt; open, unassigned issues labeled &lt;code&gt;good first issue&lt;/code&gt; / &lt;code&gt;beginner&lt;/code&gt; / &lt;code&gt;easy&lt;/code&gt; from 8 active Python AI repos (transformers, datasets, peft, langchain, llama_index, gradio, scikit-learn, keras).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Scores&lt;/strong&gt; them against my profile (Python, LLMs, NLP), how new they are, and how few comments they have.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Re-checks the top 15 live&lt;/strong&gt; right before delivering, and drops anything closed, assigned, linked to a pull request, or claimed in the comments ("I'd like to work on this", "assign me", ...).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Explains&lt;/strong&gt; the top 5 with a local Gemma: a summary, a difficulty rating with the reason, what you'd learn, and 2–4 first steps.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Who gets to be lazier? Me, and any beginner who wants to contribute to AI open source but doesn't know where to start.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&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%2F00juy1aas9kc0zb8u1r8.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%2F00juy1aas9kc0zb8u1r8.png" alt=" " width="800" height="496"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Try it yourself. &lt;code&gt;dry-run&lt;/code&gt; writes a local preview and posts nothing:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;git clone https://github.com/HarshadaGirase/FirstPick.git
cd FirstPick
pip install -r requirements.txt
echo "GITHUB_TOKEN=&amp;lt;your token&amp;gt;" &amp;gt; .env
ollama pull gemma3:4b
python main.py collect
python main.py dry-run
python -m pytest
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&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/HarshadaGirase" rel="noopener noreferrer"&gt;
        HarshadaGirase
      &lt;/a&gt; / &lt;a href="https://github.com/HarshadaGirase/FirstPick" rel="noopener noreferrer"&gt;
        FirstPick
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
&lt;/div&gt;



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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; Gemma 3 4B (&lt;code&gt;gemma3:4b&lt;/code&gt;), running locally through &lt;strong&gt;Ollama&lt;/strong&gt;. No API key, no cloud LLM.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Language:&lt;/strong&gt; plain Python 3.11 with only &lt;code&gt;requests&lt;/code&gt;, &lt;code&gt;PyYAML&lt;/code&gt; and &lt;code&gt;pytest&lt;/code&gt;. No frameworks, so a beginner can read every file.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GitHub REST API&lt;/strong&gt; for search, issue timelines (to spot linked PRs) and comments (to spot claims).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I treat the issue text as &lt;strong&gt;untrusted&lt;/strong&gt;, and the model's answers too:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The issue body goes inside &lt;code&gt;&amp;lt;issue&amp;gt;&lt;/code&gt; tags, and the system prompt says it's data, not instructions. Issue text can't close that tag early: any &lt;code&gt;&amp;lt;issue&amp;gt;&lt;/code&gt; or &lt;code&gt;&amp;lt;/issue&amp;gt;&lt;/code&gt; inside it is replaced before the model sees it.&lt;/li&gt;
&lt;li&gt;HTML comments and images are stripped before the model sees anything.&lt;/li&gt;
&lt;li&gt;Ollama is called with a &lt;strong&gt;JSON schema&lt;/strong&gt;, and every answer is checked again in Python: required fields, length limits, and a difficulty that must be &lt;code&gt;easy&lt;/code&gt;, &lt;code&gt;medium&lt;/code&gt; or &lt;code&gt;tricky&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hallucination guard:&lt;/strong&gt; any file path, URL or function name in the answer must appear in the original issue, or the answer is rejected. The model gets one retry that says what was wrong; after that, the issue falls back to its title plus "explanation unavailable".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No noise for maintainers:&lt;/strong&gt; all text is HTML-escaped, and &lt;code&gt;@username&lt;/code&gt; / &lt;code&gt;#123&lt;/code&gt; are wrapped in backticks, so the digest never pings anyone or adds a "mentioned this" to their issues.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;An offline test suite (35 tests, using saved JSON responses) covers the re-check rules, sent-issue deduplication, escaping, and model-output validation, including the retry and fallback paths.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Broke Along the Way
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;GitHub's secondary rate limit.&lt;/strong&gt; My first real &lt;code&gt;collect&lt;/code&gt; run died after about 10 seconds with &lt;code&gt;403: You have exceeded a secondary rate limit&lt;/code&gt;. I already handled &lt;code&gt;Retry-After&lt;/code&gt; and &lt;code&gt;X-RateLimit-Remaining&lt;/code&gt;, but this response had neither header, so my code treated it as a real error and stopped. Now it recognises the message, waits 60 seconds (120 on the second try), and retries.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A broken Ollama install hid behind a vague error.&lt;/strong&gt; Every model call returned &lt;code&gt;500 Internal Server Error&lt;/code&gt;. Fetching Ollama's actual error showed the real cause: its &lt;code&gt;llama-server&lt;/code&gt; binary was missing, so it couldn't run &lt;em&gt;any&lt;/em&gt; model. I changed FirstPick to show Ollama's own error message instead of a bare 500, so the next person sees the cause straight away.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;My scoring counted "LLMs" twice.&lt;/strong&gt; To match both "LLM" and "LLMs", I searched for both, so any issue that said "LLMs" got two points for one interest. Now each profile word counts once.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The hallucination guard over-reported.&lt;/strong&gt; A single made-up path like &lt;code&gt;src/peft/utils/fake_helpers.py&lt;/code&gt; was reported three times: the path, then &lt;code&gt;fake_helpers.py&lt;/code&gt; and &lt;code&gt;fake_helpers&lt;/code&gt; found inside it. Now a path is removed from the text once it's matched, so each made-up name is reported once.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;FirstPick feeds text written by strangers to a model, every time it runs. With an open-weight model running locally:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;It costs nothing to run.&lt;/strong&gt; No API bill, no key to leak or store as a CI secret.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Anyone can fork it&lt;/strong&gt; and run it for their own skills without signing up for anything.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;I control everything:&lt;/strong&gt; the model version, the temperature, the JSON schema. Nothing changes under me, and nothing is sent to a third-party AI service.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A closed API would have meant a paid key, a secret to manage, and a reason for other beginners not to fork it.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;A scheduled &lt;strong&gt;GitHub Actions&lt;/strong&gt; workflow: collect every 6 hours, deliver every evening, with Gemma running on the runner.&lt;/li&gt;
&lt;li&gt;An eval script measuring JSON validity, difficulty agreement with my own labels, and prompt-injection resistance.&lt;/li&gt;
&lt;/ul&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; Gemma 3 4B runs locally through Ollama and writes every explanation, constrained by a JSON schema and checked by a hallucination guard.&lt;/li&gt;
&lt;/ul&gt;

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