<?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: Arpita Singh</title>
    <description>The latest articles on DEV Community by Arpita Singh (@arpitasi1gh).</description>
    <link>https://dev.to/arpitasi1gh</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%2F4158263%2Fb5948960-3874-48aa-bda0-5e3603be87f9.png</url>
      <title>DEV Community: Arpita Singh</title>
      <link>https://dev.to/arpitasi1gh</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/arpitasi1gh"/>
    <language>en</language>
    <item>
      <title>Repo Necromancer: I Built an AI That Writes Eulogies for Dead GitHub Repos</title>
      <dc:creator>Arpita Singh</dc:creator>
      <pubDate>Sun, 01 Nov 2026 06:55:34 +0000</pubDate>
      <link>https://dev.to/arpitasi1gh/repo-necromancer-i-built-an-ai-that-writes-eulogies-for-dead-github-repos-7cm</link>
      <guid>https://dev.to/arpitasi1gh/repo-necromancer-i-built-an-ai-that-writes-eulogies-for-dead-github-repos-7cm</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;Every developer has a graveyard of abandoned side projects. &lt;/p&gt;

&lt;p&gt;Repos with names like &lt;code&gt;todo-app-final-v2&lt;/code&gt;, dormant since 2022, holding half-finished auth systems and a single &lt;code&gt;TODO: fix this&lt;/code&gt; comment. We pour hours into them, hit a wall, and then abandon them in the dark, never to be opened again.&lt;/p&gt;

&lt;p&gt;I built &lt;strong&gt;Repo Necromancer&lt;/strong&gt; to give these codebases a voice.&lt;/p&gt;

&lt;p&gt;You paste the URL of a dead GitHub repository. The app uses the GitHub API to read its description, how long it's been dead, and its final few commit messages. Then, it feeds that data into &lt;strong&gt;Gemma&lt;/strong&gt; and generates a dramatic, melancholic eulogy spoken from the perspective of the project itself.&lt;/p&gt;

&lt;p&gt;It laments the features it never got, the bugs it still carries, and the developer who abandoned it for a shiny new framework. &lt;/p&gt;

&lt;p&gt;&lt;em&gt;Example output from an abandoned project:&lt;/em&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"I was supposed to be a revolution. In my final days, your commits changed from 'feat: add authentication' to 'fix typo' to 'give up for tonight'. I still have that memory leak in the user controller. You left me for a Rust tutorial. I hope it makes you happy."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

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

&lt;p&gt;🔗 &lt;strong&gt;Live Demo:&lt;/strong&gt; &lt;a href="https://repo-necromancer.onrender.com" rel="noopener noreferrer"&gt;https://repo-necromancer.onrender.com&lt;/a&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
  &lt;tbody&gt;
&lt;tr&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 01&lt;/b&gt;&lt;/p&gt;
      &lt;img width="799" height="478" alt="DEMO IMAGE 01" 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%2Fbl3m73aa8ugkcmp1cgb0.png"&gt;
    &lt;/td&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 02&lt;/b&gt;&lt;/p&gt;
      &lt;img width="800" height="481" alt="DEMO IMAGE 02" 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%2F5pupbyyizd1tdiftd6ku.png"&gt;
    &lt;/td&gt;
  &lt;/tr&gt;
  &lt;tr&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 03&lt;/b&gt;&lt;/p&gt;
      &lt;img width="800" height="480" alt="DEMO IMAGE 03" 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%2F900b88ygxrmxfelafiao.png"&gt;
    &lt;/td&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 04&lt;/b&gt;&lt;/p&gt;
      &lt;img width="799" height="475" alt="DEMO IMAGE 04" 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%2Fd2epnm5c8huvllb95myp.png"&gt;
    &lt;/td&gt;
  &lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

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

&lt;p&gt;📊 &lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/arpitasi1gh/hacktoberfest/tree/main/repo-necromancer" rel="noopener noreferrer"&gt;https://github.com/arpitasi1gh/hacktoberfest/tree/main/repo-necromancer&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Stack:&lt;/strong&gt; FastAPI, HTMX, Jinja2, GitHub API, and Gemma.&lt;/p&gt;

&lt;p&gt;The architecture is split into modular services: &lt;code&gt;github_service.py&lt;/code&gt; handles data extraction, &lt;code&gt;ai_service.py&lt;/code&gt; handles the LLM logic and fallbacks, and &lt;code&gt;prompts.py&lt;/code&gt; holds the strict instructions.&lt;/p&gt;

&lt;p&gt;The hardest part was making the AI reliable. I ran into two massive bugs:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. The Monorepo Hallucination&lt;/strong&gt;&lt;br&gt;
When I first tested this on a large repository, the AI completely ignored a sub-folder and hallucinated fake folders. Because a flat list of 300 file paths is too much for an LLM to parse accurately, it just guessed.&lt;br&gt;
&lt;em&gt;The fix:&lt;/em&gt; I wrote a Python script to extract the exact top-level folders from the GitHub API tree. I then injected this into the prompt as a strict checklist: &lt;code&gt;MANDATORY FOLDERS TO ANALYZE: ...&lt;/code&gt;. The AI immediately stopped hallucinating and analyzed every single folder.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. The Thinking Model's Inner Monologue&lt;/strong&gt;&lt;br&gt;
Because I was using Gemma via Google AI Studio, I discovered it is a "thinking" model. Instead of just returning the JSON, it would output its internal planning: &lt;em&gt;"Wait, the prompt says no meta-text. Let's try formatting it like this..."&lt;/em&gt; My UI would just dump this raw text onto the screen.&lt;br&gt;
&lt;em&gt;The fix:&lt;/em&gt; I had to write a bulletproof Python parser that scans the AI's output bottom-up, finds the actual JSON object &lt;code&gt;{...}&lt;/code&gt;, and strips away the markdown and internal monologue.&lt;/p&gt;

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

&lt;p&gt;When you point a tool at a GitHub repository, you might be pointing it at a private company repo, a proprietary client project, or your own embarrassing early code. &lt;/p&gt;

&lt;p&gt;Sending that entire file tree and commit history to a closed API like OpenAI means sending your company's private codebase to a third-party server. &lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Repo Necromancer is local-first.&lt;/strong&gt; If you run it with Ollama (&lt;code&gt;gemma3:4b&lt;/code&gt;), the GitHub data, the prompt, and the analysis never leave your machine. The entire forensic autopsy happens on your laptop.&lt;/p&gt;

&lt;p&gt;To make the live demo work (since Render's free tier can't run a 4GB model), I built a hybrid architecture: it tries the local Ollama server first, and falls back to the Google AI Studio API only if needed.&lt;/p&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 via local Ollama, with a Google AI Studio Gemma fallback.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render&lt;/strong&gt; — FastAPI app deployed on Render using the cloud inference fallback.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>I built Inbox Sloth: The Local-First AI Ghostwriter for Your Most Stressful Emails</title>
      <dc:creator>Arpita Singh</dc:creator>
      <pubDate>Mon, 26 Oct 2026 06:52:50 +0000</pubDate>
      <link>https://dev.to/arpitasi1gh/i-built-inbox-sloth-the-local-first-ai-ghostwriter-for-your-most-stressful-emails-478n</link>
      <guid>https://dev.to/arpitasi1gh/i-built-inbox-sloth-the-local-first-ai-ghostwriter-for-your-most-stressful-emails-478n</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;We all have that one email we've been avoiding. The passive-aggressive client, the boss asking for a weekend update, the vendor asking for a meeting that could have been an email. Writing a reply takes 5 minutes. Avoiding it takes 3 days of low-grade anxiety.&lt;/p&gt;

&lt;p&gt;ChatGPT doesn't fix this. If you ask it to write an email, it spits out generic corporate fluff that sounds like a robot ("I hope this email finds you well...").&lt;/p&gt;

&lt;p&gt;I built &lt;strong&gt;Inbox Sloth&lt;/strong&gt; to fix that. You paste the incoming email, and it generates three distinct options:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Short:&lt;/strong&gt; A 1-sentence reply to end the thread.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Polite:&lt;/strong&gt; A standard professional response.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pushback:&lt;/strong&gt; A firm decline or boundary-setting.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;But the real magic is the &lt;strong&gt;Tone Profile&lt;/strong&gt;. You can paste a past email you wrote. The app saves it locally. Next time you generate a reply, it uses your past writing as context so the drafts actually sound like &lt;em&gt;you&lt;/em&gt;.&lt;/p&gt;

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

&lt;p&gt;🔗 &lt;strong&gt;Live Demo:&lt;/strong&gt; &lt;a href="https://inbox-sloth.onrender.com" rel="noopener noreferrer"&gt;https://inbox-sloth.onrender.com&lt;/a&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
  &lt;tbody&gt;
&lt;tr&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 01&lt;/b&gt;&lt;/p&gt;
      &lt;img width="800" height="482" alt="Demo Image 01" 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%2F4vcytevsq8ymeid2hbc3.png"&gt;
    &lt;/td&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 02&lt;/b&gt;&lt;/p&gt;
      &lt;img width="799" height="479" alt="Demo Image 02" 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%2Fw9ygaz3lizgetp2e1m5c.png"&gt;
    &lt;/td&gt;
  &lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

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

&lt;p&gt;📊 &lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/arpitasi1gh/hacktoberfest/tree/main/inbox-sloth" rel="noopener noreferrer"&gt;https://github.com/arpitasi1gh/hacktoberfest/tree/main/inbox-sloth&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Stack:&lt;/strong&gt; FastAPI, HTMX, Jinja2, and Gemma 3.&lt;/p&gt;

&lt;p&gt;I wanted this to be zero-friction. There is no login, no file uploads, no complex UI. You paste text and hit a button. HTMX handles the asynchronous loading states without needing React.&lt;/p&gt;

&lt;p&gt;The hardest part wasn't the frontend, it was the prompt engineering. To make the AI act as a ghostwriter, the backend reads the user's past emails and injects them into the prompt as "Few-Shot Examples." Then, I force the model to return a strict JSON object with exactly three keys (&lt;code&gt;short&lt;/code&gt;, &lt;code&gt;polite&lt;/code&gt;, &lt;code&gt;pushback&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;I ran into two major bugs during this build:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;The UI broke:&lt;/strong&gt; When I pasted a massive email with a very long URL, the URL had no spaces to wrap on, so it pushed the textarea outside the card. I had to add &lt;code&gt;word-wrap: break-word&lt;/code&gt; and &lt;code&gt;box-sizing: border-box&lt;/code&gt; to the CSS to fix it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The AI wrapped its JSON in markdown:&lt;/strong&gt; I had to write a bulletproof Python parser that strips the backticks and extracts only the valid JSON block from the model's output.&lt;/li&gt;
&lt;/ol&gt;

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

&lt;p&gt;Emails contain sensitive company data, client names, and sometimes confidential legal discussions. Sending this data to a closed API like OpenAI means sending your company's inbox to a third-party server.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Inbox Sloth is local-first.&lt;/strong&gt; If you run it with Ollama (&lt;code&gt;gemma3:4b&lt;/code&gt;), your emails never leave your laptop. The entire app is a private, closed loop.&lt;/p&gt;

&lt;p&gt;To make the app demoable as a live link (since free hosting can't run a 4GB model), I built a hybrid fallback: it tries the local Ollama server first, and falls back to the Google AI Studio API only if needed. This local-first setup is only possible because open models are finally small enough to run on consumer hardware.&lt;/p&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 via local Ollama, with a Google AI Studio Gemma fallback.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render&lt;/strong&gt; — FastAPI app deployed on Render using the cloud inference fallback.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>week3</category>
    </item>
    <item>
      <title>I Built Rehearse to Sell A Voice-First, Local-First AI Interview Coach</title>
      <dc:creator>Arpita Singh</dc:creator>
      <pubDate>Sun, 18 Oct 2026 10:29:10 +0000</pubDate>
      <link>https://dev.to/arpitasi1gh/i-built-rehearse-to-sell-a-voice-first-local-first-ai-interview-coach-33c9</link>
      <guid>https://dev.to/arpitasi1gh/i-built-rehearse-to-sell-a-voice-first-local-first-ai-interview-coach-33c9</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;I built &lt;strong&gt;Rehearse&lt;/strong&gt;, a voice-first and local-first AI interviewer.&lt;/p&gt;

&lt;p&gt;You know how you read a list of "Top 50 Interview Questions," think "I could answer that," and then completely freeze when you actually have to speak out loud? That is the exact problem this solves.&lt;/p&gt;

&lt;p&gt;ChatGPT doesn't help here. If you ask it to practice with you, it gives you a giant essay, says "Great job!", and never pushes back. A real interview is a conversation. Rehearse actually talks to you (using voice), cuts you off if you ramble, and forces you to answer the hard follow-up questions. At the end, it grades you out of 10 and tells you how a better candidate would have answered.&lt;/p&gt;

&lt;p&gt;I didn't want this to just be for software engineers. You can type "Marketing," "Nursing," or "Sales," and paste a real job description (I tested it with a Deloitte JD), and the AI changes its tone and questions to match that field. And if you pick "Stress Test" mode, it acts like a genuinely hostile interviewer.&lt;/p&gt;

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

&lt;p&gt;🔗 &lt;strong&gt;Live Demo:&lt;/strong&gt; &lt;a href="https://rehearse-g3q4.onrender.com" rel="noopener noreferrer"&gt;https://rehearse-g3q4.onrender.com&lt;/a&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
  &lt;tbody&gt;
&lt;tr&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 01&lt;/b&gt;&lt;/p&gt;
      &lt;img width="799" height="479" alt="Demo Image 01" 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%2Fl0x800af656l4esei8th.png"&gt;
    &lt;/td&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 02&lt;/b&gt;&lt;/p&gt;
      &lt;img width="800" height="454" alt="Demo Image 02" 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%2Fw2w9s10h9yx0k4576978.png"&gt;
    &lt;/td&gt;
  &lt;/tr&gt;
  &lt;tr&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 03&lt;/b&gt;&lt;/p&gt;
      &lt;img width="800" height="481" alt="Demo Image 03" 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%2Fpcq3gu9sgxs9irmkk46j.png"&gt;
    &lt;/td&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 04&lt;/b&gt;&lt;/p&gt;
      &lt;img width="800" height="471" alt="Demo Image 04" 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%2Fg30zw3l59d53rl47vhzw.png"&gt;
    &lt;/td&gt;
  &lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

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

&lt;p&gt;📊 &lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/arpitasi1gh/hacktoberfest/tree/main/rehearse" rel="noopener noreferrer"&gt;https://github.com/arpitasi1gh/hacktoberfest/tree/main/rehearse&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;The stack is FastAPI, HTMX, and Gemma. I used the browser's built-in microphone and text-to-speech so it works instantly without saving audio files to a server.&lt;/p&gt;

&lt;p&gt;The hardest part wasn't the code, it was stopping the AI from being too nice. I had to write a very strict set of rules for the prompt. It is explicitly banned from saying things like "Great answer" or "I understand." If you say "we built a system," it is instructed to interrupt you and ask, "What did you specifically do?"&lt;/p&gt;

&lt;p&gt;Then I ran into two hilarious bugs. &lt;/p&gt;

&lt;p&gt;First, Gemma on Google's API is a "thinking" model. Instead of just asking the question, it would write out its entire internal thought process: &lt;em&gt;"Wait, the prompt says no meta-text. Let's try asking this..."&lt;/em&gt; Because my app reads the text out loud, it was literally speaking the AI's internal self-doubt to the user. &lt;/p&gt;

&lt;p&gt;Second, when I pasted a massive job description into the context field, the local model got confused and started wrapping its JSON output in markdown backticks (&lt;br&gt;
&lt;br&gt;
&lt;code&gt;json {...}&lt;/code&gt;&lt;br&gt;
&lt;br&gt;
). My UI just dumped the raw JSON onto the screen. I had to write a bulletproof Python parser that scans the AI's output bottom-up, ignores the monologue, strips the markdown, and grabs only the final &lt;code&gt;{"question": "..."}&lt;/code&gt; object.&lt;/p&gt;

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

&lt;p&gt;Interviewing is vulnerable. You are anxious, alone, and talking to a screen. If this used a closed API, it would mean sending your awkward voice recordings to a giant tech company's server.&lt;/p&gt;

&lt;p&gt;Because it uses open-source AI, this app can run 100% locally. If you run it with Ollama, nothing leaves your laptop.&lt;/p&gt;

&lt;p&gt;To make the app demoable as a live link (since free hosting can't run a 4GB model), I built a fallback: it tries the local Ollama server first, and if you don't have it running, it uses the Google API. This local-first setup is only possible because open models are finally small enough to run on normal laptops.&lt;/p&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 via local Ollama, with a Google AI Studio Gemma fallback.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render&lt;/strong&gt; — FastAPI app deployed on Render using the cloud inference fallback.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
    </item>
    <item>
      <title>I built Greenstep – An Open-Source Web App That Picks Your Next Walk</title>
      <dc:creator>Arpita Singh</dc:creator>
      <pubDate>Mon, 12 Oct 2026 06:29:04 +0000</pubDate>
      <link>https://dev.to/arpitasi1gh/i-built-greenstep-an-open-source-app-that-picks-your-next-walk-ke3</link>
      <guid>https://dev.to/arpitasi1gh/i-built-greenstep-an-open-source-app-that-picks-your-next-walk-ke3</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;I always struggle with the typical "touch grass" advice. It's true, but when you're exhausted after a long day of classes, you don't just magically decide to go outside. There's no clear signal that &lt;em&gt;this&lt;/em&gt; is the right moment, and no specific place to go. So you just stay inside and scroll.&lt;/p&gt;

&lt;p&gt;I built &lt;strong&gt;Greenstep&lt;/strong&gt; to fix that. It’s a small web app that tells you exactly when and where to go outside. &lt;/p&gt;

&lt;p&gt;You give it four inputs: your city, how much you slept, your screen time, and your mood. The app geocodes your city, fetches the live weather, queries Google Maps to find a real cafe or park nearby, and then uses Gemma 3 to write one specific action for you. &lt;/p&gt;

&lt;p&gt;It doesn't give you a list of options. It gives you one short, finishable action. &lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Walk to ROOTED GROUNDS CAFE and order a black coffee.&lt;/strong&gt;&lt;br&gt;
&lt;em&gt;Mild weather and a middling mood. A change of environment is appropriate.&lt;/em&gt;&lt;br&gt;
&lt;strong&gt;First step:&lt;/strong&gt; Grab your jacket.&lt;br&gt;
&lt;strong&gt;Timebox:&lt;/strong&gt; 25 min.&lt;br&gt;
&lt;strong&gt;Fallback:&lt;/strong&gt; Stand outside your building for 10 minutes.&lt;/p&gt;
&lt;/blockquote&gt;

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

&lt;p&gt;🔗 &lt;strong&gt;Live Demo:&lt;/strong&gt; &lt;a href="https://hacktoberfest-m24i.onrender.com" rel="noopener noreferrer"&gt;https://hacktoberfest-m24i.onrender.com&lt;/a&gt;&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fgithub.com%2Fuser-attachments%2Fassets%2F7888c130-a6a0-42a9-a9fc-bc898be4b4ef" class="article-body-image-wrapper"&gt;&lt;img width="2940" height="1762" alt="Screenshot 2026-11-11 at 1 06 35 AM" src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fgithub.com%2Fuser-attachments%2Fassets%2F7888c130-a6a0-42a9-a9fc-bc898be4b4ef"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
  &lt;tbody&gt;
&lt;tr&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 01&lt;/b&gt;&lt;/p&gt;
      &lt;img width="2940" height="1762" alt="Screenshot 2026-11-11 at 1 06 35 AM" src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fgithub.com%2Fuser-attachments%2Fassets%2F7888c130-a6a0-42a9-a9fc-bc898be4b4ef"&gt;
    &lt;/td&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 02&lt;/b&gt;&lt;/p&gt;
      &lt;img width="2822" height="1700" alt="Screenshot 2026-11-11 at 1 06 09 AM" src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fgithub.com%2Fuser-attachments%2Fassets%2F55d0ea4a-3a30-4871-b058-618fcbb2e0e8"&gt;
    &lt;/td&gt;
  &lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

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

&lt;p&gt;📊 &lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/arpitasi1gh/hacktoberfest/tree/main/greenstep" rel="noopener noreferrer"&gt;https://github.com/arpitasi1gh/hacktoberfest/tree/main/greenstep&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;The stack is FastAPI, HTMX, and SQLite. &lt;/p&gt;

&lt;p&gt;The pipeline uses free APIs to get context (Nominatim for geocoding, Open-Meteo for weather, SerpApi for finding local places), and then feeds that context into Gemma 3 to generate the final action. &lt;/p&gt;

&lt;p&gt;Getting a local AI model to actually work reliably was the hardest part of this project. &lt;/p&gt;

&lt;p&gt;I started with &lt;code&gt;gemma3:4b&lt;/code&gt; running locally through Ollama. It kept hallucinating weird, poetic phrases like "a walk might be conducive to stillness." I had to stop writing nice instructions and literally ban words in the system prompt. I added a rule saying &lt;code&gt;NEVER use these words: might, could, perhaps, simply, just, opportunity, wellness, conducive, stillness&lt;/code&gt;. &lt;/p&gt;

&lt;p&gt;Once the tone was fixed, the JSON broke. The model kept wrapping its output in markdown backticks, which crashed my Python script. I ended up writing a custom parser that strips the markdown, cuts the preamble before the first &lt;code&gt;{&lt;/code&gt;, and cuts the trailing text after the last &lt;code&gt;}&lt;/code&gt;. &lt;/p&gt;

&lt;p&gt;Finally, because people's sleep and screen time data is pretty personal, I wanted the app to run locally. However, Render's free tier can't host a 4GB model. So I added a simple fallback. If you run it locally, it uses Ollama. If you deploy it, you switch one environment variable to use Google AI Studio's Gemma endpoint. Same prompt, same output, just a different runtime.&lt;/p&gt;

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

&lt;p&gt;If I built this with a closed API, I'd be sending someone's daily mood and sleep schedule to a third-party server. I'd also be paying per prediction. &lt;/p&gt;

&lt;p&gt;Because I used open models (Gemma via Ollama), the app can run completely offline. Your behavioral data never leaves your laptop. You don't pay a cent for inference, and you can actually read the prompt in the repo to see exactly what the AI is doing. &lt;/p&gt;

&lt;p&gt;Open models make it possible to build personal tools that respect privacy without breaking the bank.&lt;/p&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 via Ollama (local) and Gemma via Google AI Studio (deployed).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of SerpApi&lt;/strong&gt;: Finding real cafes and parks instead of hallucinating generic places.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render&lt;/strong&gt;: Deployed FastAPI app using the Gemma cloud fallback.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
    </item>
    <item>
      <title>What Now? – A Local-First AI Next-Action Coach for a Tired College Student</title>
      <dc:creator>Arpita Singh</dc:creator>
      <pubDate>Mon, 05 Oct 2026 06:58:30 +0000</pubDate>
      <link>https://dev.to/arpitasi1gh/what-now-a-local-first-ai-next-action-coach-for-a-tired-college-student-2ec6</link>
      <guid>https://dev.to/arpitasi1gh/what-now-a-local-first-ai-next-action-coach-for-a-tired-college-student-2ec6</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;What Now?&lt;/strong&gt; for a friend of mine. She's 20, in her second year of CS, and she has the kind of schedule that looks fine on paper and falls apart by evening.&lt;/p&gt;

&lt;p&gt;Here's what I kept noticing. She isn't confused about what matters. She has a DBMS assignment due tomorrow, placement prep she's behind on, a gym membership she pays for and doesn't use, and a real need to rest. She knows all of it. But at 7:30 PM, after a full day of college, the gap between knowing and starting is where the evening goes. One reel becomes ten. "I'll start after this episode" becomes three. The assignment stays open in a tab she doesn't touch.&lt;/p&gt;

&lt;p&gt;It isn't laziness. Deciding what to start takes more energy than starting it.&lt;/p&gt;

&lt;p&gt;What she needed wasn't another planner. She's tried those. They hand her a list of twelve things and call it clarity, which is the same problem wearing a nicer font. What she needed was for something else to make the call. One thing. Right now. Small enough that saying yes doesn't feel like a commitment.&lt;/p&gt;

&lt;p&gt;That's what I built. A loop, not an app. She opens it, sets her energy on a 1–5 slider, says how many minutes she has, and hits one button. The app pulls her active tasks, sends the context to a language model running locally, and gets back a single action:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;Open the DBMS assignment PDF and write the first SQL query for Q1.&lt;/em&gt;&lt;br&gt;
&lt;em&gt;Why: due tomorrow at 9 AM, high urgency.&lt;/em&gt;&lt;br&gt;
&lt;em&gt;First step: open the PDF. Timebox: 15 minutes.&lt;/em&gt;&lt;br&gt;
&lt;em&gt;Fallback: if still low energy, review one DSA problem instead.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That's the whole output. No list. No options. No priority matrix asking her to rank things she's too tired to rank.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The flow, end to end:&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;She adds her tasks once, whenever she has the energy to think about them. Title, deadline, roughly how long it'll take, how much energy it costs, and how much it matters.&lt;/li&gt;
&lt;li&gt;Later that evening, she opens the app. Nothing to re-enter. Her tasks are already there.&lt;/li&gt;
&lt;li&gt;She sets energy to 2 and minutes to 45. The check-in form is the first thing on the page, because that's the only decision I want her making at 7:30 PM.&lt;/li&gt;
&lt;li&gt;The app sends her context to Gemma 3 4B, running locally through Ollama, and gets back one action with a reason, a two-minute first step, a timebox, and a fallback.&lt;/li&gt;
&lt;li&gt;The action card appears below the form. She reads it, and either starts or she doesn't.&lt;/li&gt;
&lt;li&gt;Either way, she taps &lt;strong&gt;Done&lt;/strong&gt;, &lt;strong&gt;Skip&lt;/strong&gt;, or &lt;strong&gt;Blocked&lt;/strong&gt;. The outcome gets logged. If she hits Done, the task's progress slider increments by 10%. &lt;/li&gt;
&lt;li&gt;When a task hits 100% progress, the UI automatically grays it out and strikes it through, marking it complete. The loop runs again next time.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The goal was never to turn her into a productivity machine. She can still watch her shows and play guitar. The goal was to close the gap between knowing and starting.&lt;/p&gt;

&lt;p&gt;The app is designed to run locally on her laptop alongside Ollama. However, because I wanted to share a live link, I built a cloud fallback. If Ollama is not running locally, the app automatically switches to the Google AI Studio API.&lt;/p&gt;

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

&lt;p&gt;🔗 &lt;strong&gt;Live Demo:&lt;/strong&gt; &lt;a href="https://what-now-g8eg.onrender.com" rel="noopener noreferrer"&gt;https://what-now-g8eg.onrender.com&lt;/a&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
  &lt;tbody&gt;
&lt;tr&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 01&lt;/b&gt;&lt;/p&gt;
      &lt;img width="800" height="478" alt="Demo Image 01" 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%2F3ea9faog6oxwd6wx2mhk.png"&gt;
    &lt;/td&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 02&lt;/b&gt;&lt;/p&gt;
      &lt;img width="799" height="496" alt="Demo Image 02" 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%2F0r5buu2xni5k61r9sosm.png"&gt;
    &lt;/td&gt;
  &lt;/tr&gt;
  &lt;tr&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 03&lt;/b&gt;&lt;/p&gt;
      &lt;img width="800" height="520" alt="Demo Image 03" 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%2Ftgge3cqyrkxno90m7r1r.png"&gt;
    &lt;/td&gt;
    &lt;td&gt;
      &lt;p&gt;&lt;b&gt;Demo Image 04&lt;/b&gt;&lt;/p&gt;
      &lt;img width="800" height="499" alt="Demo Image 04" 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%2Flxc2n67nmkuh5dtl1jmp.png"&gt;
    &lt;/td&gt;
  &lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

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

&lt;p&gt;📊 &lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/arpitasi1gh/hacktoberfest/tree/main/what-now" rel="noopener noreferrer"&gt;https://github.com/arpitasi1gh/hacktoberfest/tree/main/what-now&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Stack:&lt;/strong&gt; FastAPI with HTMX on the front end, SQLite for storage, and Gemma 3 4B via Ollama.&lt;/p&gt;

&lt;p&gt;Four files carry the project:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;prompts.py&lt;/code&gt; holds the strict system prompt. It forces the model to return JSON with exactly five fields, and explicitly tells it to pick ONE task based on deadline and energy.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;ai_service.py&lt;/code&gt; handles the inference. It tries Ollama first using &lt;code&gt;format: "json"&lt;/code&gt;, but catches connection errors and falls back to Gemini if the local server is offline.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;db.py&lt;/code&gt; handles the SQLite schema, including tasks, progress, and action history.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;main.py&lt;/code&gt; wires the routes and uses HTMX Out-Of-Band (OOB) swaps. When a task is completed, HTMX updates the task list, the "Open" count badge, and the disabled state of the generate button all in a single response.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The prompt does most of the heavy lifting. It tells the model to respect energy levels (if energy is 2 or below, no high-energy tasks unless the deadline is under 24 hours), to keep timeboxes under 25 minutes, and to write like a peer rather than a coach. No "let's crush it" or "you got this."&lt;/p&gt;

&lt;p&gt;Swapping models is one environment variable:&lt;br&gt;
&lt;code&gt;OLLAMA_MODEL=qwen2.5:7b uv run uvicorn src.what_now.main:app&lt;/code&gt;&lt;/p&gt;

&lt;p&gt;I went with Ollama over the alternatives because it handles quantization, memory, and model downloads without me writing any of that myself. It exposes a plain HTTP endpoint, so the Python side is a single &lt;code&gt;httpx&lt;/code&gt; call.&lt;/p&gt;

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

&lt;p&gt;Her task list contains her exam schedule, her placement prep, and goals she hasn't shared with anyone. A closed API would mean sending that to a server, paying per call, and trusting a vendor's logs.&lt;/p&gt;

&lt;p&gt;Running Gemma locally through Ollama means:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Nothing leaves her laptop.&lt;/li&gt;
&lt;li&gt;No API bills and no rate limits.&lt;/li&gt;
&lt;li&gt;Every prompt is in the repo. She can read exactly what the model sees.&lt;/li&gt;
&lt;li&gt;When a better open model comes out, she swaps one environment variable.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That last one is the part I keep coming back to. If I'd built this on a closed API, the project would be a wrapper around someone else's product, and it would stop working the day they change their pricing or deprecate the model. Because the model runs on her machine, it's hers. She can read it, change it, or replace it without asking anyone.&lt;/p&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 via Ollama (local) and Gemma via Google AI Studio (deployed fallback).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render&lt;/strong&gt; — FastAPI app deployed on Render using the cloud inference fallback.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>weekendchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Introduction to Object-Oriented Programming (OOP)</title>
      <dc:creator>Arpita Singh</dc:creator>
      <pubDate>Fri, 02 Oct 2026 18:38:53 +0000</pubDate>
      <link>https://dev.to/arpitasi1gh/introduction-to-object-oriented-programming-oop-2g9l</link>
      <guid>https://dev.to/arpitasi1gh/introduction-to-object-oriented-programming-oop-2g9l</guid>
      <description>&lt;p&gt;Object-Oriented Programming (OOP) is a programming paradigm centered around structuring code into self-contained units called objects. Instead of storing related data across separate, disconnected arrays, OOP bundles data and actions together to model real-world entities cleanly.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key Building Blocks:&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;i) Class: The structural blueprint or template (e.g., User or BankAccount). It defines what fields and methods its instances will have.&lt;/p&gt;

&lt;p&gt;ii) Object: A concrete instance created from a class residing in memory (e.g., user1).&lt;/p&gt;

&lt;p&gt;iii) State (Fields): The data or properties an object holds, such as email, balance, or status.&lt;/p&gt;

&lt;p&gt;iv) Behavior (Methods): The actions an object can perform, such as login(), deposit(), or withdraw().&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The 4 Core Pillars:&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;i) Encapsulation: Hides internal data using private fields and exposes controlled access via getters and setters to protect data integrity.&lt;/p&gt;

&lt;p&gt;ii) Abstraction: Exposes essential features while hiding underlying implementation details using abstract classes and interfaces.&lt;/p&gt;

&lt;p&gt;iii) Inheritance: Enables child classes to reuse properties and methods from a parent class via extends.&lt;/p&gt;

&lt;p&gt;iv) Polymorphism: Allows methods to execute differently based on the object invoking them, through method overloading and dynamic overriding.&lt;/p&gt;

&lt;p&gt;By enforcing separation of concerns, OOP prevents silent data corruption, eliminates synchronization bugs, and makes large software applications modular, scalable, and easy to maintain.&lt;/p&gt;

</description>
      <category>productivity</category>
      <category>oop</category>
    </item>
  </channel>
</rss>
