<?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: Parth Kulkarni</title>
    <description>The latest articles on DEV Community by Parth Kulkarni (@parth_kulkarni_f56a656693).</description>
    <link>https://dev.to/parth_kulkarni_f56a656693</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%2F3889208%2F394cdd2c-2f65-41f7-a9c6-42e97f365e36.jpg</url>
      <title>DEV Community: Parth Kulkarni</title>
      <link>https://dev.to/parth_kulkarni_f56a656693</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/parth_kulkarni_f56a656693"/>
    <language>en</language>
    <item>
      <title>bhoot.sh: I Put the Ghost of a Frustrated Indian Senior Dev on My Desktop (100% Local with Gemma 2)</title>
      <dc:creator>Parth Kulkarni</dc:creator>
      <pubDate>Wed, 28 Oct 2026 13:03:57 +0000</pubDate>
      <link>https://dev.to/parth_kulkarni_f56a656693/bhootsh-i-put-the-ghost-of-a-frustrated-indian-senior-dev-on-my-desktop-100-local-with-gemma-2-17i2</link>
      <guid>https://dev.to/parth_kulkarni_f56a656693/bhootsh-i-put-the-ghost-of-a-frustrated-indian-senior-dev-on-my-desktop-100-local-with-gemma-2-17i2</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;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr7uofyxl754igvypxlkh.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr7uofyxl754igvypxlkh.gif" alt="Bhoot-Sh" width="600" height="338"&gt;&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;When I looked at the &lt;strong&gt;"Ghost in the Machine"&lt;/strong&gt; prompt, I noticed almost everyone was building haunted CLI shells or digital memory altars. As an Indian developer, my mind immediately went to a different kind of horror: &lt;strong&gt;the evolutions of Indian ghost cinema—and the terrifying spirit of a fed-up senior developer watching you copy-paste from Stack Overflow at 2:00 AM.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Classic Indian horror evolved from melodramatic spirits in ancestral &lt;em&gt;havelis&lt;/em&gt; (&lt;em&gt;Mahal&lt;/em&gt;, 1949) and Ramsay Brothers folklore into modern, self-aware "Good Ghosts" (&lt;em&gt;Stree&lt;/em&gt;, &lt;em&gt;Bhoot&lt;/em&gt;) who haunt you with lived realities rather than cheap jump scares. Simultaneously, I was inspired by the floating interactive "Loony" lobby companion from &lt;em&gt;Brawl Stars&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;So I combined the two and built &lt;strong&gt;&lt;code&gt;bhoot.sh&lt;/code&gt;&lt;/strong&gt; (&lt;strong&gt;"Bhoot"&lt;/strong&gt; = Ghost in Hindi/Marathi): a &lt;strong&gt;frameless, 60 FPS transparent desktop wraith&lt;/strong&gt; that literally possesses your Windows OS, stalks your IDE peripheral vision, tracks your mouse cursor with glowing hollow eyes, and roasts your bad coding habits out loud in &lt;strong&gt;Romanized Hinglish&lt;/strong&gt; and &lt;strong&gt;English&lt;/strong&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  What &lt;code&gt;bhoot.sh&lt;/code&gt; Actually Does on Your Screen:
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Autonomous IDE Stalker &amp;amp; Real-Time Eye Tracking:&lt;/strong&gt; Unlike a passive terminal script, &lt;code&gt;bhoot.sh&lt;/code&gt; is a live &lt;code&gt;PyQt6&lt;/code&gt; desktop entity whose glowing pupils calculate trigonometric vectors (&lt;code&gt;math.hypot&lt;/code&gt;) to follow your mouse cursor across the screen in real time. Instead of blocking the code you are actively typing, it glides to targeted IDE zones:

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Stack Overflow Copy / Terrible Git Commit:&lt;/strong&gt; Hovers along the right-hand margin of VS Code, glaring sideways at your editor.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Terminal Crash / Test Failure:&lt;/strong&gt; Glides down to hover directly over your bottom terminal panel, shaking with a crimson aura.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;API Key Leak / Reels Distraction:&lt;/strong&gt; Creeps right up next to your mouse cursor to confront you.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;3-Tier Sin Escalation (&lt;code&gt;short&lt;/code&gt; → &lt;code&gt;medium&lt;/code&gt; → &lt;code&gt;long&lt;/code&gt;):&lt;/strong&gt; Every coding sin you commit is counted. Your first mistake gets a &lt;code&gt;short&lt;/code&gt; warning (&lt;em&gt;"Ctrl C, Ctrl V. Wah, engineer."&lt;/em&gt;). Your second gets a &lt;code&gt;medium&lt;/code&gt; roast (&lt;em&gt;"Stack Overflow wale bhaiya ko salary de de, project toh unhone hi likha hai."&lt;/em&gt;). Hit the &lt;code&gt;long&lt;/code&gt; tier (3rd+ offense) and the wraith becomes &lt;strong&gt;ENRAGED&lt;/strong&gt;—triggering a full-screen, click-through blood-crimson vignette (&lt;code&gt;VoidDimmer&lt;/code&gt;) across your entire monitor while local &lt;strong&gt;Gemma 2 (&lt;code&gt;gemma2:2b&lt;/code&gt;)&lt;/strong&gt; improvises a fresh roast.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Poltergeist Mode (Anti-Poke Retaliation):&lt;/strong&gt; If you get annoyed and spam-click the ghost 5 times in under 3 seconds, it retaliates. It triggers a harmless UI-only &lt;strong&gt;Poltergeist Override&lt;/strong&gt;, dimming your screen blood-red, rapidly teleporting across 6 chaotic screen coordinates so you can't catch it, and flashing a fake &lt;code&gt;System32&lt;/code&gt; / Git branch deletion warning before laughing at you.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;p&gt;&lt;strong&gt;🌐 Live Cloud Telemetry &amp;amp; Browser Voice Playground:&lt;/strong&gt; &lt;a href="https://bhoot-sh.onrender.com" rel="noopener noreferrer"&gt;https://bhoot-sh.onrender.com&lt;/a&gt; &lt;em&gt;(Explore live MongoDB sin logs and test ElevenLabs voice roasts directly in your browser!)&lt;/em&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  📸 Screenshots
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;The Grim Wraith Stalking VS Code (Side-Margin Glare &amp;amp; Eye-Tracking):&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fwa5gj2u2csdyqql94r7d.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%2Fwa5gj2u2csdyqql94r7d.png" alt="The Grim Wraith Stalking VS Code" width="800" height="489"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Enraged Tier + Full-Screen Crimson &lt;code&gt;VoidDimmer&lt;/code&gt; + Poltergeist Mode:&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F357u47cvofuv1w0dsxs8.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%2F357u47cvofuv1w0dsxs8.png" alt="Enraged Tier + Full-Screen Crimson  raw `VoidDimmer` endraw  + Poltergeist Mode" width="800" height="469"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Sentry Agent Tracing Spans (&lt;code&gt;ai.inference&lt;/code&gt; on Local Gemma 2):&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fl4p5rd9gq2vro03dey14.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%2Fl4p5rd9gq2vro03dey14.png" alt="Sentry Agent Tracing Spans" width="437" height="329"&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%2Fuhdfujym7vwi6ep6omlo.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%2Fuhdfujym7vwi6ep6omlo.png" alt="ai.inference" width="440" height="215"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;MongoDB Atlas Persistent Sin Ledger (&lt;code&gt;bhoot_sh.coding_sins&lt;/code&gt;):&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F19p1bzxo6zugpyd6jc9q.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%2F19p1bzxo6zugpyd6jc9q.png" alt="MongoDB Atlas Persistent Sin Ledger" width="800" height="355"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Render Live Web Dashboard &amp;amp; MongoDB Telemetry Integration:&lt;/strong&gt;&lt;/p&gt;

&lt;h2&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%2Fdgn6ehbkx9prx42m0sxk.png" alt="Render Live Web Dashboard" width="800" height="382"&gt;
&lt;/h2&gt;

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


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/p-rth12999" rel="noopener noreferrer"&gt;
        p-rth12999
      &lt;/a&gt; / &lt;a href="https://github.com/p-rth12999/bhoot-sh" rel="noopener noreferrer"&gt;
        bhoot-sh
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      A floating, offline desktop ghost that haunts your screen, tracks your coding sins, and roasts you in Hinglish using local Gemma 2.
    &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;👻 bhoot.sh&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;A reincarnated, highly judgmental senior developer haunting your desktop. Built for the &lt;strong&gt;Hacktoberfest 2026: Ghost in the Machine&lt;/strong&gt; challenge, &lt;strong&gt;bhoot.sh&lt;/strong&gt; is a frameless, transparent desktop wraith that monitors your clipboard and active windows 100% locally—roasting your bad coding habits in &lt;strong&gt;Hinglish&lt;/strong&gt; and &lt;strong&gt;English&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;If you copy-paste from Stack Overflow for the tenth time, accidentally copy a live API key, open Instagram Reels near a deadline, or push &lt;code&gt;"fix bug"&lt;/code&gt; straight to &lt;code&gt;main&lt;/code&gt;, this phantom glides across your screen, locks its glowing eyes onto your mouse cursor, dims your monitor blood-red, and yells at you.&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;✨ Key Features&lt;/h2&gt;
&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Autonomous Stalker Roaming &amp;amp; Eye-Tracking:&lt;/strong&gt; Glides dynamically to specific IDE zones without blocking your active typing line—hovering along the right editor margin when you copy-paste, flying down to glare at your VS Code terminal on a crash, or creeping up beside your cursor when you open distraction tabs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;5-Mood&lt;/strong&gt;…&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/p-rth12999/bhoot-sh" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;





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

&lt;p&gt;I designed &lt;code&gt;bhoot.sh&lt;/code&gt; around a &lt;strong&gt;hybrid local-first architecture&lt;/strong&gt; so that the UI remains butter-smooth at 60 FPS while monitoring system-level events without ever freezing or leaking private clipboard data to the cloud.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;+-----------------------------------------------------------------------+
|                         WINDOWS DESKTOP / IDE                         |
|  [Clipboard (pyperclip)]   [Window Titles (pygetwindow)]   [Cursor]   |
+-------------+----------------------------+--------------------+-------+
              |                            |                    |
              v                            v                    |
+-------------------------------------------------------+       |
|          src/watcher.py (Background QThread)          |       |
|  • Regex Secret Scanner (sk-*, ghp_*, mongodb+srv://) |       |
|  • Distraction Tab Detector (Reels, YouTube, X)       |       |
|  • Terminal Crash &amp;amp; Console.log Detector              |       |
+---------------------------+---------------------------+       |
                            |                                   |
                            v                                   |
+-------------------------------------------------------+       |
|         src/memory.py (MongoDB Atlas + Local)         |       |
|  • Logs sin category &amp;amp; timestamp to bhoot_sh.sins     |       |
|  • Escalates severity: 1=SHORT -&amp;gt; 2=MEDIUM -&amp;gt; 3+=LONG |       |
+---------------------------+---------------------------+       |
                            |                                   |
                            v                                   |
+-------------------------------------------------------+       |
|     src/ghost_brain.py + GemmaWorker (Async QThread)  |       |
|  • Tier 1-3: Instant lookup from scenariohindi.json   |       |
|    &amp;amp; scenarioenglish.json + Pre-baked ElevenLabs MP3  |       |
|  • Tier 4+: Few-Shot Prompting on Local Gemma 2 (2B)  |       |
|    via Ollama, wrapped in Sentry Agent Tracing Spans  |       |
+---------------------------+---------------------------+       |
                            |                                   |
                            v                                   v
+-----------------------------------------------------------------------+
|              src/ui_overlay.py (PyQt6 Frameless Overlay)              |
|  • 60FPS Procedural Grim Wraith + Rising Smoke Particles              |
|  • Real-Time Pupil Tracking + QPropertyAnimation Zone Roaming         |
|  • Full-Screen Click-Through VoidDimmer + Native winmm.dll Audio      |
+-----------------------------------------------------------------------+
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;h3&gt;
  
  
  The Stack Breakdown:
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Open-Weight Local Model (&lt;code&gt;Gemma 2 2B&lt;/code&gt; via Ollama):&lt;/strong&gt;&lt;br&gt;
Running inside an asynchronous &lt;code&gt;GemmaWorker(QThread)&lt;/code&gt; in &lt;code&gt;src/main.py&lt;/code&gt;, &lt;code&gt;gemma2:2b&lt;/code&gt; powers our dynamic roast engine. Instead of generating generic AI slop, &lt;code&gt;src/ghost_brain.py&lt;/code&gt; dynamically injects lines from our bilingual datasets (&lt;code&gt;scenariohindi.json&lt;/code&gt; and &lt;code&gt;scenarioenglish.json&lt;/code&gt;) as &lt;strong&gt;few-shot style examples&lt;/strong&gt; into the prompt so Gemma nails the exact sarcastic Hinglish cadence (&lt;em&gt;"Bas kar bhai"&lt;/em&gt;, &lt;em&gt;"Deploy karleeee"&lt;/em&gt;, &lt;em&gt;"Mera past life yaad aa gaya"&lt;/em&gt;).&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;System Telemetry Daemon (&lt;code&gt;src/watcher.py&lt;/code&gt;):&lt;/strong&gt;&lt;br&gt;
A dedicated &lt;code&gt;QThread&lt;/code&gt; monitors clipboard changes (&lt;code&gt;pyperclip&lt;/code&gt;), active window titles (&lt;code&gt;pygetwindow&lt;/code&gt;), and system clock hours. It uses regex pattern matching to classify 18+ distinct developer sins—from exposed &lt;code&gt;sk-&lt;/code&gt; / &lt;code&gt;ghp_&lt;/code&gt; / &lt;code&gt;mongodb+srv://&lt;/code&gt; keys and &lt;code&gt;git push --force&lt;/code&gt; commands to late-night 2:00 AM coding sessions.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Dual-Voice Pre-Baked Audio Engine (&lt;code&gt;src/tts_baker.py&lt;/code&gt; + &lt;code&gt;winmm.dll&lt;/code&gt;):&lt;/strong&gt;&lt;br&gt;
We built a custom baker script using &lt;strong&gt;ElevenLabs&lt;/strong&gt; (&lt;code&gt;eleven_multilingual_v2&lt;/code&gt;) with two distinct personas: &lt;strong&gt;Daniel&lt;/strong&gt; for thick, dramatic Hinglish delivery and &lt;strong&gt;Clyde&lt;/strong&gt; for gravelly English roasts. By dropping &lt;code&gt;stability&lt;/code&gt; to &lt;code&gt;0.25&lt;/code&gt; and cranking &lt;code&gt;style&lt;/code&gt; to &lt;code&gt;0.85&lt;/code&gt; with injected pause punctuation (&lt;code&gt;...&lt;/code&gt; and &lt;code&gt;—&lt;/code&gt;), the voice sighs, rasps, and mocks naturally. At runtime, &lt;code&gt;ui_overlay.py&lt;/code&gt; plays these &lt;code&gt;.mp3&lt;/code&gt; files directly through Windows' native Multimedia Control Interface (&lt;code&gt;ctypes.windll.winmm.mciSendStringW&lt;/code&gt;) with dynamic alias rotation—achieving &lt;strong&gt;0ms network latency&lt;/strong&gt; and &lt;strong&gt;100% offline execution&lt;/strong&gt;.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Persistent Sin Memory (&lt;code&gt;MongoDB Atlas&lt;/code&gt;):&lt;/strong&gt;&lt;br&gt;
&lt;code&gt;src/memory.py&lt;/code&gt; connects to our MongoDB Atlas cluster (&lt;code&gt;bhoot_sh.coding_sins&lt;/code&gt;) to store an immutable ledger of every coding sin, severity tier, and timestamp (with automatic offline fallback if Wi-Fi is off).&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Agent Observability (&lt;code&gt;Sentry Agent Tracing&lt;/code&gt;):&lt;/strong&gt;&lt;br&gt;
Every call to local &lt;code&gt;gemma2:2b&lt;/code&gt; is wrapped in a &lt;code&gt;sentry_sdk.start_span(op="ai.inference")&lt;/code&gt; transaction tagged with &lt;code&gt;language&lt;/code&gt;, &lt;code&gt;severity_tier&lt;/code&gt;, and &lt;code&gt;model&lt;/code&gt;, giving us full visibility into local CPU/GPU inference latency.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;


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

&lt;p&gt;Building &lt;code&gt;bhoot.sh&lt;/code&gt; exposed two massive reasons why &lt;strong&gt;open-weight models and local inference are non-negotiable&lt;/strong&gt; for OS-level developer tools:&lt;/p&gt;
&lt;h3&gt;
  
  
  1. The Clipboard Privacy Paradox
&lt;/h3&gt;

&lt;p&gt;For &lt;code&gt;bhoot.sh&lt;/code&gt; to catch you copying code from Stack Overflow or accidentally copying a production &lt;code&gt;.env&lt;/code&gt; secret (&lt;code&gt;api_key_exposed&lt;/code&gt;), &lt;strong&gt;it has to inspect your system clipboard and active window titles every second&lt;/strong&gt;.&lt;br&gt;
If &lt;code&gt;bhoot.sh&lt;/code&gt; were built on a closed cloud API, watching for leaked API keys would literally require &lt;strong&gt;sending your leaked API keys and proprietary source code over the internet to a third-party server&lt;/strong&gt;! That turns a fun developer companion into a massive security vulnerability. Because &lt;code&gt;gemma2:2b&lt;/code&gt; runs 100% locally via Ollama on my own laptop, my clipboard never leaves &lt;code&gt;localhost&lt;/code&gt;. You can turn off your Wi-Fi completely, and the ghost still haunts your machine.&lt;/p&gt;
&lt;h3&gt;
  
  
  2. Escaping the Cloud Quota &amp;amp; Safety-Filter Wall (Architectural Lesson &amp;amp; Future Scope)
&lt;/h3&gt;

&lt;p&gt;First, closed commercial LLMs often neuter "roast" or "horror" personas with overly sanitized corporate refusals. Open-weight &lt;code&gt;gemma2:2b&lt;/code&gt; let us use few-shot Hinglish prompting to capture raw, authentic Indian developer banter without censorship.&lt;/p&gt;

&lt;p&gt;Second, during development, we hit a classic closed-API bottleneck with voice synthesis: a desktop agent that reacts to every clipboard event triggers dozens of times an hour. Calling a cloud TTS API dynamically on every error introduces 2+ seconds of lag and burns through free-tier character quotas (and paywalled voice libraries) in a single afternoon. &lt;br&gt;
While our hybrid solution (&lt;code&gt;tts_baker.py&lt;/code&gt; pre-baking core ElevenLabs MP3s + live Gemma 2 text) solved this for v1.0, it proved why &lt;strong&gt;our immediate v2.0 roadmap is integrating a 100% open-weight local neural TTS (&lt;code&gt;Kokoro-82M&lt;/code&gt; / &lt;code&gt;Parler-TTS&lt;/code&gt;)&lt;/strong&gt; alongside Ollama—allowing every unscripted Gemma 2 roast to be voiced locally on-device with zero API limits forever.&lt;/p&gt;


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


&lt;div class="ltag-agent-session"&gt;
  &lt;div class="agent-session-header"&gt;
    
    &lt;span class="agent-session-tool-icon-badge" title="GitHub Copilot"&gt;
&lt;/span&gt;
    &lt;span class="agent-session-title"&gt;bhoot.sh smoke test&lt;/span&gt;
  &lt;/div&gt;

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

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Run a quick session testing bhoot.sh and provide a public DevRelay session link or Liquid embed tag.&lt;/p&gt;

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

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Inspected the project entry point and README runtime command, then ran a local Python bytecode compilation smoke test for the src package.&lt;/p&gt;

                &lt;/div&gt;
              &lt;div class="agent-session-tool-call"&gt;
                
                  &lt;span class="agent-session-chevron"&gt;▸&lt;/span&gt;
                  &lt;span class="tool-name tool-name-powershell"&gt;powershell&lt;/span&gt;
                    &lt;span class="tool-input-preview"&gt;python -m compileall -q src&lt;/span&gt;
                
                &lt;div class="agent-session-tool-detail"&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Input&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;python -m compileall -q src&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Output&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;PASS: Python bytecode compilation completed for src&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                &lt;/div&gt;
              &lt;/div&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;The smoke test passed. Submitting this sanitized session transcript to DevRelay.&lt;/p&gt;

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

  &lt;div class="agent-session-footer"&gt;
    &lt;span class="agent-session-meta"&gt;
        2 of 2 messages
    &lt;/span&gt;
  &lt;/div&gt;
&lt;/div&gt;






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

&lt;p&gt;The moment &lt;code&gt;bhoot.sh&lt;/code&gt; truly landed for me was at &lt;strong&gt;2:00 AM&lt;/strong&gt; while debugging &lt;code&gt;ui_overlay.py&lt;/code&gt;. &lt;/p&gt;

&lt;p&gt;Initially, when I triggered a live Gemma 2 roast, the synchronous inference call blocked the main thread and froze the ghost's animation—making it feel like a buggy script. After moving Gemma 2 into a background &lt;code&gt;QThread&lt;/code&gt; and replacing our early cartoon balloon prototype with the hooded dark wraith, I accidentally copied a stack trace three times in a row while testing. &lt;/p&gt;

&lt;p&gt;Suddenly, my entire monitor flashed blood-crimson (&lt;code&gt;VoidDimmer&lt;/code&gt;), the hooded wraith glided straight down over my VS Code terminal with its red pupils locked onto my cursor, and the speakers rasped out in Hindi:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"Deploy karleeee, bol raha tha na? Ab rollback kar, aur apni izzat bhi saath mein rollback kar le."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Then I spam-clicked it in frustration, accidentally triggering &lt;strong&gt;Poltergeist Mode&lt;/strong&gt;, and watched it violently glitch and teleport across my screen yelling about &lt;code&gt;System32&lt;/code&gt;. It crossed the line from "a Python script I'm writing" to "an annoying, hilarious roommate living inside my laptop."&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; Uses local &lt;code&gt;gemma2:2b&lt;/code&gt; via Ollama inside an asynchronous &lt;code&gt;QThread&lt;/code&gt; with bilingual few-shot prompting (&lt;code&gt;scenariohindi.json&lt;/code&gt; &amp;amp; &lt;code&gt;scenarioenglish.json&lt;/code&gt;) to improvise context-aware developer roasts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of ElevenLabs:&lt;/strong&gt; Uses &lt;code&gt;eleven_multilingual_v2&lt;/code&gt; with custom low-stability/high-style emotional voice settings across dual personas (&lt;strong&gt;Daniel&lt;/strong&gt; for Hinglish, &lt;strong&gt;Clyde&lt;/strong&gt; for English) via &lt;code&gt;src/tts_baker.py&lt;/code&gt; to voice the desktop ghost.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of MongoDB Atlas:&lt;/strong&gt; Uses &lt;code&gt;pymongo&lt;/code&gt; in &lt;code&gt;src/memory.py&lt;/code&gt; to persist developer coding sins, timestamps, and escalation tiers (&lt;code&gt;short&lt;/code&gt; / &lt;code&gt;medium&lt;/code&gt; / &lt;code&gt;long&lt;/code&gt;) to a cloud &lt;code&gt;bhoot_sh.coding_sins&lt;/code&gt; collection.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Sentry Agent Tracing:&lt;/strong&gt; Instruments local &lt;code&gt;gemma2:2b&lt;/code&gt; inference calls in &lt;code&gt;src/ghost_brain.py&lt;/code&gt; with &lt;code&gt;sentry_sdk.start_span(op="ai.inference")&lt;/code&gt; to track latency and metadata across languages and escalation tiers.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render:&lt;/strong&gt; Deployed the live &lt;code&gt;bhoot.sh&lt;/code&gt; Cloud Telemetry Dashboard &amp;amp; Browser Voice Playground on Render (&lt;a href="https://bhoot-sh.onrender.com" rel="noopener noreferrer"&gt;https://bhoot-sh.onrender.com&lt;/a&gt;), connecting live to MongoDB Atlas for real-time sin logging and remote audio playback.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>SlothCloset</title>
      <dc:creator>Parth Kulkarni</dc:creator>
      <pubDate>Thu, 22 Oct 2026 12:41:49 +0000</pubDate>
      <link>https://dev.to/parth_kulkarni_f56a656693/slothcloset-2iod</link>
      <guid>https://dev.to/parth_kulkarni_f56a656693/slothcloset-2iod</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;Every morning starts with the exact same low-level friction: standing half-awake in front of an open wardrobe, suffering from choice paralysis, and burning mental energy trying to figure out if an olive cargo pant clashes with a navy jersey tee or if your favorite hoodie is still in the laundry basket. &lt;/p&gt;

&lt;p&gt;&lt;strong&gt;SlothCloset Studio&lt;/strong&gt; (&lt;code&gt;SLOTHCLOSET_OS&lt;/code&gt;) is an offline-first, open-weight wardrobe deck-builder that treats your closet like a &lt;em&gt;Clash Royale&lt;/em&gt; deck-building constraint graph. Designed for anyone who wants to automate their daily routine and be unapologetically lazy before their morning coffee, SlothCloset eliminates morning decision fatigue in under a millisecond:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;🃏 Weekly Wash-Cycle Rotation Solver (3–7 Days):&lt;/strong&gt; Computes a complete week of non-overlapping outfits in &lt;code&gt;~0.5ms&lt;/code&gt; to &lt;code&gt;2.0ms&lt;/code&gt;. It automatically enforces a configurable &lt;strong&gt;1–3 day laundry/rest cooldown&lt;/strong&gt; on tops and bottoms while allowing daily staples like watches and sneakers to rotate naturally.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;🔒 Standalone Outfit Synthesizer (Lock / Ban Veto):&lt;/strong&gt; Want to wear one specific hoodie today and let math figure out the rest? Lock any garment in place (or veto items you aren't in the mood for), choose a meta style archetype (&lt;em&gt;Standard Meta&lt;/em&gt;, &lt;em&gt;Monochrome Beatdown&lt;/em&gt;, &lt;em&gt;Earth-Tone Cycle&lt;/em&gt;, or &lt;em&gt;High-Contrast Aggro&lt;/em&gt;), and synthesize high-synergy combinations with built-in anti-clone rotation.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;🎨 Multi-Color Swatch &amp;amp; Cutout Pipeline:&lt;/strong&gt; Supports multi-color garments (up to 3–4 HEX swatches per piece), strips messy bedroom backgrounds locally using open-weight &lt;strong&gt;&lt;code&gt;rembg&lt;/code&gt; (U2Net)&lt;/strong&gt;, or fetches clean studio flat-lay catalog twins via &lt;strong&gt;SerpApi&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;🧠 Tactical AI Stylist &amp;amp; "5th-Card Unlock":&lt;/strong&gt; Powered by &lt;strong&gt;Google Gemma 2 (&lt;code&gt;gemma2:2b&lt;/code&gt;)&lt;/strong&gt;, an optional non-blocking AI analyst writes concise synergy breakdowns for any drafted outfit and recommends a single missing wardrobe piece to unlock new combinations—complete with a live product visual pulled via SerpApi.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;p&gt;🚀 &lt;strong&gt;Live Cloud Deployment (24/7):&lt;/strong&gt; &lt;a href="https://sloth-closet12999.streamlit.app/" rel="noopener noreferrer"&gt;https://sloth-closet12999.streamlit.app/&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Studio UI Walkthrough (&lt;code&gt;SLOTHCLOSET_OS&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;00. Minimalist Landing Index:&lt;/strong&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%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F59ealu3x93hv8i9eqrjw.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%2F59ealu3x93hv8i9eqrjw.png" alt="Landing Page" width="799" height="469"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;01. Outfit Planner &amp;amp; Archetype Engine:&lt;/strong&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%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fq9d28scap7k7mlawrjw3.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%2Fq9d28scap7k7mlawrjw3.png" alt="Outfit Planner" width="800" height="463"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Standalone Outfit Builder (Force-Lock &amp;amp; Ban/Veto Controls):&lt;/strong&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%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fnodwqipm1isk9pk1zk44.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%2Fnodwqipm1isk9pk1zk44.png" alt="Outfit Builder (Lock/BAN)" width="800" height="461"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Weekly Rotation Matrix (3–7 Day Wash-Cycle Solver):&lt;/strong&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%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyl8vefbra3lz2sc23u07.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%2Fyl8vefbra3lz2sc23u07.png" alt="Weekly Builder" width="800" height="569"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;02. Wardrobe Asset Repository (Clean / Laundry / Skip States):&lt;/strong&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%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5axc5yb9svx88te26yrf.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%2F5axc5yb9svx88te26yrf.png" alt="Wardrobe Assets" width="800" height="569"&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/p-rth12999" rel="noopener noreferrer"&gt;
        p-rth12999
      &lt;/a&gt; / &lt;a href="https://github.com/p-rth12999/Sloth-CLoset" rel="noopener noreferrer"&gt;
        Sloth-CLoset
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      An offline-first, open-weight wardrobe deck builder that automates your daily outfit choices so you never have to stand in front of your closet stressing out again.
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;p&gt;🎯 The Problem
Every morning comes with the exact same friction: staring into a dark wardrobe, suffering from choice paralysis, and wasting mental energy trying to figure out what pants go with what shirt. Existing wardrobe apps are bloated, invasive, and require sending personal photos of your clothes and room to third-party cloud servers.&lt;/p&gt;
&lt;p&gt;💡 The Solution: SlothCloset Studio
Inspired by Clash Royale deck-building systems and RoyaleAPI asset management, SlothCloset treats your wardrobe like a card game deck. It lets you build a Weekly Outfit Rotation in under a millisecond using offline heuristic math, features a Ban/Skip List Veto, handles Wash-Cycle Cooldowns, and leverages Google Gemma 2 (via Ollama) for deep local AI styling advice.&lt;/p&gt;
&lt;p&gt;✨ Key Features
🃏 RoyaleAPI-Style Weekly Rotation: Automatically plans 3 to 7 days of unique non-overlapping outfits at once.&lt;/p&gt;
&lt;p&gt;⚡ Sub-Millisecond Python Heuristic Graph: Zero-LLM latency for the core color-matching engine (~0.5ms compute time).&lt;/p&gt;
&lt;p&gt;🧺…&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/p-rth12999/Sloth-CLoset" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;&lt;em&gt;(Fallback Repository Link: &lt;a href="https://github.com/p-rth12999/Sloth-CLoset" rel="noopener noreferrer"&gt;https://github.com/p-rth12999/Sloth-CLoset&lt;/a&gt;)&lt;/em&gt;&lt;/p&gt;




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

&lt;p&gt;Early in development, I realized a major trap with AI-first apps: if you put an LLM directly in the critical path of generating a 7-day outfit rotation, the user has to stare at a loading spinner for 10+ seconds just to pick a t-shirt. That isn't "lazy"—that's annoying.&lt;/p&gt;

&lt;p&gt;Instead, I architected SlothCloset around a &lt;strong&gt;Hybrid Heuristic + Local Open-Weight AI&lt;/strong&gt; pipeline:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Sub-Millisecond Python Constraint Graph (&lt;code&gt;0.001s&lt;/code&gt; Core):&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;Every clothing card stores up to 3 dominant HEX codes alongside its category (&lt;code&gt;Top&lt;/code&gt;, &lt;code&gt;Layer&lt;/code&gt;, &lt;code&gt;Bottom&lt;/code&gt;, &lt;code&gt;Footwear&lt;/code&gt;, &lt;code&gt;Accessory&lt;/code&gt;), laundry state, and ban status.&lt;/li&gt;
&lt;li&gt;Using RGB luminance and HSV saturation math, the engine classifies swatches into color families (&lt;code&gt;neutral_dark&lt;/code&gt;, &lt;code&gt;neutral_light&lt;/code&gt;, &lt;code&gt;neutral_grey&lt;/code&gt;, &lt;code&gt;earth&lt;/code&gt;, &lt;code&gt;warm_pop&lt;/code&gt;, &lt;code&gt;cool&lt;/code&gt;) and scores pairwise color harmony against &lt;code&gt;meta_rules.json&lt;/code&gt; archetypes.&lt;/li&gt;
&lt;li&gt;For standalone builds, an anti-clone diversity penalty tracks recently drafted items on screen so every click of &lt;strong&gt;+ Generate New Outfit&lt;/strong&gt; rotates fresh combinations without repeating the same jacket or jeans.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Open-Weight Local AI (&lt;code&gt;Google Gemma 2&lt;/code&gt; via Ollama &amp;amp; &lt;code&gt;rembg&lt;/code&gt; U2Net):&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Google Gemma 2 (&lt;code&gt;gemma2:2b&lt;/code&gt; via Ollama):&lt;/strong&gt; Rather than blocking the UI, Gemma 2 acts as an on-demand Tactical Stylist. When triggered on any outfit card, it evaluates the drafted combination in structured JSON mode, explains &lt;em&gt;why&lt;/em&gt; the color contrast works, and identifies a "Missing Asset" that would bridge the user's current closet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Open-Weight Vision (&lt;code&gt;rembg&lt;/code&gt; / U2Net + Pillow Mediancut):&lt;/strong&gt; When uploading a raw photo of a garment locally, &lt;code&gt;rembg&lt;/code&gt; strips the background entirely on-device, and Pillow quantizes the remaining opaque pixels to automatically extract the garment's top 3 HEX color swatches.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dual-Runtime Frontend &amp;amp; Data Layer:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Local Studio SPA:&lt;/strong&gt; Built with &lt;strong&gt;FastAPI&lt;/strong&gt;, &lt;strong&gt;Tailwind CSS&lt;/strong&gt;, and &lt;strong&gt;Alpine.js&lt;/strong&gt; in a stark dark-mode developer aesthetic inspired by Andrej Karpathy's minimalist tooling and Superdesign.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cloud Twin:&lt;/strong&gt; Wrapped the exact same heuristic engine and database layer into &lt;strong&gt;Streamlit&lt;/strong&gt; so anyone can test the solver 24/7 in the browser.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Persistence &amp;amp; Visuals:&lt;/strong&gt; Backed by &lt;strong&gt;MongoDB Atlas&lt;/strong&gt; (with automatic local JSON fallback for 100% offline resilience) and &lt;strong&gt;SerpApi Google Images&lt;/strong&gt; to pull clean flat-lay catalog visuals for wardrobe items and Gemma 2's shopping recommendations.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ol&gt;




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

&lt;p&gt;Your wardrobe—and photos of your bedroom, laundry, and daily habits—is deeply personal data. Existing "smart closet" apps are bloated SaaS subscriptions that force you to upload personal photos to proprietary cloud buckets and pay per API call just to organize your own clothes.&lt;/p&gt;

&lt;p&gt;Open innovation made three critical things possible for SlothCloset that a closed API couldn't:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Zero-Telemetry Privacy:&lt;/strong&gt; By combining local &lt;strong&gt;Ollama (&lt;code&gt;gemma2:2b&lt;/code&gt;)&lt;/strong&gt; with open-weight &lt;strong&gt;&lt;code&gt;rembg&lt;/code&gt; (U2Net)&lt;/strong&gt;, photos of your clothes and your daily styling queries never have to leave &lt;code&gt;localhost&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero-Cost Infinite Experimentation:&lt;/strong&gt; A student or indie developer shouldn't have to worry about token billing every time they re-roll their weekly outfit schedule. Open weights let you run tactical styling analysis as many times as you want for $0.00.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hackable Rulesets:&lt;/strong&gt; Because the color-harmony matrix (&lt;code&gt;meta_rules.json&lt;/code&gt;) and constraint solver are completely transparent Python code rather than an opaque black-box prompt, anyone can fork the repo and tune the weights for their own local climate, uniform rules, or personal aesthetic.&lt;/li&gt;
&lt;/ol&gt;




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

&lt;p&gt;Let's do the lazy math:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Spending &lt;strong&gt;4 minutes every morning&lt;/strong&gt; debating what to wear—or realizing your go-to shirt is still in the wash—costs nearly &lt;strong&gt;30 minutes a week&lt;/strong&gt; (over &lt;strong&gt;24 hours a year&lt;/strong&gt;) of pure decision fatigue before 8:00 AM.&lt;/li&gt;
&lt;li&gt;With &lt;strong&gt;SlothCloset Studio&lt;/strong&gt;, I spend &lt;strong&gt;10 seconds on Sunday night&lt;/strong&gt; clicking &lt;code&gt;Generate Weekly Plan&lt;/code&gt; (or &lt;code&gt;1.08 ms&lt;/code&gt; locking in a single shirt on a random weekday morning) to get a pre-validated, color-matched rotation that already respects my laundry cycle.&lt;/li&gt;
&lt;li&gt;I used it all week while building this submission: locking in my Colorblock Zip Hoodie for a chilly morning commute and letting the engine automatically pair it with Olive Relaxed Cargos and Court Sneakers without me having to think twice.&lt;/li&gt;
&lt;/ul&gt;




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


&lt;div class="ltag-agent-session"&gt;
  &lt;div class="agent-session-header"&gt;
    
    &lt;span class="agent-session-tool-icon-badge" title="GitHub Copilot"&gt;
&lt;/span&gt;
    &lt;span class="agent-session-title"&gt;SlothCloset Studio: deterministic outfit synthesis and Gemma stylist&lt;/span&gt;
  &lt;/div&gt;

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

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Inspect SlothCloset Studio's main.py, streamlit_app.py, and meta_rules.json. Summarize how the sub-millisecond Python color-harmony heuristic engine, wash-cycle cooldown solver, and local Ollama Gemma 2 stylist work together.&lt;/p&gt;

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

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Inspected the three requested artifacts without transmitting secrets or private machine paths. The project has a deterministic scoring layer, a cooldown-aware weekly planner, and an optional local Ollama enrichment layer.&lt;/p&gt;

                &lt;/div&gt;
              &lt;div class="agent-session-tool-call"&gt;
                
                  &lt;span class="agent-session-chevron"&gt;▸&lt;/span&gt;
                  &lt;span class="tool-name tool-name-workspaceinspection"&gt;workspace_inspection&lt;/span&gt;
                    &lt;span class="tool-input-preview"&gt;main.py; streamlit_app.py; meta_rules.json&lt;/span&gt;
                
                &lt;div class="agent-session-tool-detail"&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Input&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;main.py; streamlit_app.py; meta_rules.json&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Output&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;Found color-family classification, pairwise harmony scoring, archetype modifiers, standalone outfit enumeration, weekly availability/cooldown scheduling, UI timing instrumentation, and an Ollama Gemma 2 JSON analysis endpoint.&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                &lt;/div&gt;
              &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;The heuristic engine classifies HEX colors into neutral_dark, neutral_light, neutral_grey, earth, warm_pop, or cool. It averages configured pairwise family scores from meta_rules.json, applies archetype-specific boosts/penalties and vibe bonuses, adds a deterministic ID tie-breaker, and clamps the result. Standalone generation enumerates valid category combinations, honors locked/excluded items, ranks by harmony minus recent-item repetition penalties, and returns the best distinct build.&lt;/p&gt;

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

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;The weekly solver receives only clean, unbanned items, builds a 3–7 day plan, and tracks last-worn day plus wear count. Tops obey the selected wash/dry cooldown; bottoms and layers require more than one day rest; footwear and accessories can recur daily. Each day's candidate combinations are scored by the same harmony engine and adjusted for freshness, with deterministic fallback picks when the eligible pool is exhausted.&lt;/p&gt;

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

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;The local Ollama Gemma 2 stylist is an optional second-stage explainer, not the constraint solver. The app sends the selected cards and archetype to localhost Ollama using a configurable local model (default gemma2:2b), requests strict JSON, and asks for a concise battle strategy plus one wardrobe unlock recommendation. The response is sanitized/parsing-guarded and paired with an optional SerpApi visual; if Ollama fails, a local deterministic fallback explanation is returned.&lt;/p&gt;

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

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;The Streamlit surface times solver calls with perf_counter and displays the plan, scores, palettes, cooldown/reuse badges, and optional Gemma analysis. meta_rules.json is the data-driven policy layer: base score, color families, pairings, and balanced/monochrome/earth/contrast archetype descriptions.&lt;/p&gt;

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

  &lt;div class="agent-session-footer"&gt;
    &lt;span class="agent-session-meta"&gt;
        6 of 6 messages
    &lt;/span&gt;
  &lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;During my agent-assisted build sessions, we iterated through three major architectural pivots for &lt;strong&gt;SlothCloset Studio&lt;/strong&gt;:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Moving the LLM out of the critical path:&lt;/strong&gt; We pivoted from slow LLM outfit generation to a deterministic &lt;code&gt;&amp;lt; 1.0ms&lt;/code&gt; Python RGB/HSV constraint solver (&lt;code&gt;score_outfit_combo&lt;/code&gt;), keeping &lt;strong&gt;Google Gemma 2 (&lt;code&gt;gemma2:2b&lt;/code&gt;)&lt;/strong&gt; as an asynchronous Tactical Stylist.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Solving the "Single-Item Clone" bottleneck:&lt;/strong&gt; When early standalone builds kept repeating the same hoodie and sneakers, we expanded the closet to 16 multi-swatch assets and implemented an anti-clone diversity penalty that tracks recently drafted items on screen.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dual-Runtime Packaging:&lt;/strong&gt; We paired the minimalist dark-mode FastAPI + Alpine.js local studio UI with a 24/7 Streamlit Cloud twin backed by MongoDB Atlas and SerpApi.&lt;/li&gt;
&lt;/ol&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; Uses Google's open-weight &lt;strong&gt;Gemma 2 (&lt;code&gt;gemma2:2b&lt;/code&gt;)&lt;/strong&gt; model via Ollama (with cloud fallback support) in structured JSON output mode to generate on-demand outfit synergy critiques and "5th-Card Unlock" wardrobe upgrade recommendations.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of MongoDB Atlas:&lt;/strong&gt; Persists multi-color wardrobe documents, live laundry states (&lt;code&gt;is_clean&lt;/code&gt;), and veto flags (&lt;code&gt;is_banned&lt;/code&gt;) in a &lt;strong&gt;MongoDB Atlas&lt;/strong&gt; cluster with seamless local JSON fallback for offline-first reliability.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of SerpApi:&lt;/strong&gt; Uses the &lt;strong&gt;SerpApi Google Images API&lt;/strong&gt; in two ways: (1) converting text descriptions of clothes into clean, isolated "Studio Twin" flat-lay product visuals for the closet grid, and (2) fetching real-time shopping visuals and inspection links for Gemma 2's recommended missing wardrobe items.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Sift</title>
      <dc:creator>Parth Kulkarni</dc:creator>
      <pubDate>Sun, 18 Oct 2026 16:50:24 +0000</pubDate>
      <link>https://dev.to/parth_kulkarni_f56a656693/sift-1m93</link>
      <guid>https://dev.to/parth_kulkarni_f56a656693/sift-1m93</guid>
      <description>&lt;h1&gt;
  
  
  Offline Hinglish DM-to-Dispatch Engine for Instagram Sellers
&lt;/h1&gt;

&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;Across India, thousands of small Instagram thrift stores, crochet artists, home bakers, and independent D2C boutiques run their entire business inside Instagram and WhatsApp DMs. &lt;/p&gt;

&lt;p&gt;There is just one massive bottleneck: &lt;strong&gt;customers don't order in clean forms—they order in messy Hinglish&lt;/strong&gt; (a blend of Hindi and English). A typical DM looks like this:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"Bhai 2 oversized black acid-wash tees bhej do L size mein. Name: Rohan Deshmukh, Ph: 9823011223. Address: Flat 402, Sai Residency, Kothrud, Pune. GPay kar diya hai 1398 rs."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Every single night, store owners spend 1.5 to 2 hours manually reading through chat dumps, copying names, phone numbers, sizes, and addresses, and pasting them into Excel sheets to generate courier dispatch labels.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Sift&lt;/strong&gt; solves this completely offline. It is a local-first AI desktop utility that takes raw, unstructured Hinglish chat dumps and extracts validated, courier-ready dispatch tables in seconds—running 100% locally on the seller's laptop at &lt;strong&gt;₹0.00 ($0.00) inference cost&lt;/strong&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  Who Would Pay For It &amp;amp; The Business Model
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Target Customers:&lt;/strong&gt; Instagram thrift pages, home bakers, small boutiques, and WhatsApp commerce sellers handling 15–100+ orders a day.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why They Pay:&lt;/strong&gt; Saving 45+ hours of manual data entry every month without paying per-message cloud API fees or risking customer privacy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pricing Model:&lt;/strong&gt; Offered as a &lt;strong&gt;₹499/month (~$6/mo)&lt;/strong&gt; flat local license or a &lt;strong&gt;₹1,999 (~$24) one-time local setup&lt;/strong&gt; where I configure Ollama, the open-weight model, and Sift directly on their machine. Because all inference runs on the user's own hardware, my cloud hosting cost is $0, making every sale 100% net profit.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Live Interactive Demo Link:&lt;/strong&gt; &lt;a href="https://sift12999.streamlit.app/" rel="noopener noreferrer"&gt;https://sift12999.streamlit.app/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Screenshot:&lt;/strong&gt; &lt;/li&gt;
&lt;/ul&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%2Fxeamdcx8jqhnfvba6qg5.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%2Fxeamdcx8jqhnfvba6qg5.png" alt="Sift Local Dashboard Screenshot" width="800" height="354"&gt;&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;a href="https://github.com/p-rth12999/Sift" rel="noopener noreferrer"&gt;https://github.com/p-rth12999/Sift&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;Sift is architected from the ground up around &lt;strong&gt;local open-weight inference&lt;/strong&gt; and &lt;strong&gt;deterministic schema enforcement&lt;/strong&gt;:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Local Inference Engine (Ollama):&lt;/strong&gt; Instead of routing messages to a paid cloud endpoint, Sift connects to a local Ollama server (&lt;code&gt;localhost:11434&lt;/code&gt;) running compact open-weight models—specifically &lt;strong&gt;Google's &lt;code&gt;gemma2:2b&lt;/code&gt;&lt;/strong&gt; and &lt;strong&gt;Meta's &lt;code&gt;llama3.2:3b&lt;/code&gt;&lt;/strong&gt;. These 2B–3B parameter models are the sweet spot for consumer laptops: they consume ~2 GB of RAM, run rapidly on standard laptop CPUs without requiring a discrete GPU, and understand Indian Hinglish slang (&lt;em&gt;"bhej do"&lt;/em&gt;, &lt;em&gt;"GPay kar diya"&lt;/em&gt;, &lt;em&gt;"COD rakhna"&lt;/em&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Strict Structured Outputs (Pydantic):&lt;/strong&gt; Small models can occasionally get chatty. To make the extraction 100% reliable for a real business workflow, I paired the Ollama Python SDK with &lt;strong&gt;Pydantic&lt;/strong&gt; (&lt;code&gt;OrderExtraction.model_json_schema()&lt;/code&gt;). By passing the JSON schema directly into the model's &lt;code&gt;format&lt;/code&gt; parameter at &lt;code&gt;temperature=0.0&lt;/code&gt;, the open-weight model is constrained at the token level to emit strictly valid JSON matching our exact database columns (&lt;code&gt;customer_name&lt;/code&gt;, &lt;code&gt;phone_number&lt;/code&gt;, &lt;code&gt;item_name&lt;/code&gt;, &lt;code&gt;quantity&lt;/code&gt;, &lt;code&gt;size&lt;/code&gt;, &lt;code&gt;total_amount&lt;/code&gt;, &lt;code&gt;payment_status&lt;/code&gt;, &lt;code&gt;delivery_address&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero-Cloud Persistence &amp;amp; Dispatch UI (SQLite3 + Streamlit + Pandas):&lt;/strong&gt; Extracted orders are committed immediately to a local SQLite database (&lt;code&gt;sift_orders.db&lt;/code&gt;). The Streamlit dashboard supports &lt;strong&gt;Batch DM Processing&lt;/strong&gt; (splitting bulk chat dumps via &lt;code&gt;---&lt;/code&gt;), live revenue and payment status KPIs, and one-click CSV/Excel dispatch sheet exports.&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  System Architecture
&lt;/h3&gt;



&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart LR
    A[📥 Messy Hinglish DMs] --&amp;gt;|Batch Input| B(Streamlit Seller UI)
    B --&amp;gt;|Raw Text| C{Local Ollama Engine\ngemma2:2b / llama3.2:3b}
    D[Pydantic Schema\nOrderExtraction] --&amp;gt;|Token-Level JSON Constraint| C
    C --&amp;gt;|Validated JSON| E[(SQLite Local DB\nsift_orders.db)]
    E --&amp;gt; F[📊 Live Revenue KPIs]
    E --&amp;gt; G[📦 Courier CSV Export]&lt;/code&gt;&lt;/pre&gt;






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

&lt;p&gt;Building Sift with closed, proprietary APIs (like GPT-4o or Claude) would have killed this side hustle before it even launched for three specific reasons:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Unit Economics That Actually Work for Micro-Businesses:&lt;/strong&gt; Indian Instagram sellers operate on razor-thin margins. If a tool charges them per token or requires a $20/month API subscription on top of the software cost, they simply will not buy it—they will go back to manual copy-pasting. Running open-weight models locally on their existing laptop means &lt;strong&gt;marginal cost per parsed order is strictly $0.00&lt;/strong&gt;, allowing me to price Sift at an impulse-buy flat rate with zero server overhead.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;100% Customer PII Sovereignty:&lt;/strong&gt; Every single order DM contains highly sensitive Personally Identifiable Information (full names, personal mobile numbers, and exact home addresses). Sending customer residential data to external third-party cloud servers creates privacy and trust barriers. With Sift's local open-weight architecture, &lt;strong&gt;customer data never leaves the seller's laptop&lt;/strong&gt;—in fact, the entire pipeline runs with Wi-Fi completely turned off.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model Swappability &amp;amp; Zero Vendor Lock-In:&lt;/strong&gt; Using Ollama as a local harness lets sellers seamlessly switch between &lt;code&gt;gemma2:2b&lt;/code&gt; and &lt;code&gt;llama3.2:3b&lt;/code&gt; depending on their laptop's RAM and CPU speed, or fine-tune a small model on regional dialects (Marathi/Hindi/Tamil transliteration) without ever worrying about API deprecations or sudden pricing changes.&lt;/li&gt;
&lt;/ol&gt;




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


&lt;div class="ltag-agent-session"&gt;
  &lt;div class="agent-session-header"&gt;
    
    &lt;span class="agent-session-tool-icon-badge" title="GitHub Copilot"&gt;
&lt;/span&gt;
    &lt;span class="agent-session-title"&gt;Sift review — Offline Hinglish DM-to-Dispatch Engine&lt;/span&gt;
  &lt;/div&gt;

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

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Review the Sift project files app.py, parser.py, and database.py for the Hacktoberfest Week 2 Side Hustle DEV Challenge, then save/share this agent session to my DEV account and provide the exact agent_session Liquid tag and URL.&lt;/p&gt;

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

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Reviewed Sift's Streamlit UI, Pydantic/Ollama parser, SQLite persistence layer, README, and requirements. Verified all three bundled fallback-parser samples: quantities, payment statuses, amounts, names, and addresses matched the sample intent. Verified all three reviewed Python files have no syntax errors.&lt;/p&gt;

                &lt;/div&gt;
                &lt;div&gt;
                  &lt;div class="agent-session-text agent-session-text-collapse"&gt;
                    &lt;p&gt;Findings: (1) High reliability risk: parser.py catches every Exception around Ollama import/call/JSON decoding and silently falls back, so missing Ollama, timeouts, malformed JSON, schema violations, and runtime defects are indistinguishable from a successful parse. Surface the failure and validate the response with OrderExtraction before persisting. (2) Data-integrity risk: OrderExtraction uses unconstrained str/int/float fields even though the prompt says payment_status is strictly Paid/COD/Pending and phone/quantity/amount have stated semantics; enforce those constraints before database insertion. (3) Fallback scope risk: cloud_demo_fallback is explicitly demo-sample-oriented (hard-coded names/items and narrow regexes), so production messages can be misclassified while appearing successful. Label it as demo-only or make the fallback fail visibly for unsupported inputs. (4) Portability risk: database.py uses a process-current-working-directory relative database path, so launching Streamlit from another directory can create/use a different database. Resolve the DB path from the application directory or configure it explicitly. (5) Transaction/resource robustness: database operations manually open/close connections without context managers or rollback handling; an exception can leave resources open and a partial batch can be reported ambiguously. (6) UX consistency: app.py advertises llama3.2:3b in the caption but defaults the selector to gemma2:2b, and always reports the local Ollama engine as active even when the fallback path will be used.&lt;/p&gt;

                  &lt;/div&gt;
                  Show more
                &lt;/div&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Static checks: Pylance reported missing imports for streamlit and ollama in the selected interpreter, while pandas and pydantic resolved. This is an environment/dependency installation issue rather than a Python syntax issue; requirements.txt declares both packages. Runtime fallback probe passed for the three bundled samples.&lt;/p&gt;

                &lt;/div&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Challenge context: Hacktoberfest Open-Source AI Challenge: Week 2, theme Side Hustle, asks for open-weight/open-source AI that someone would actually pay for. Judging emphasizes writing quality, theme relevance, creativity, technical execution, and optional partner technology. Submission requires a DEV post using the #hf26challenge tag; the challenge context says October 18 at 11:59 PM PDT, and saving/embedding an agent session is optional but helpful.&lt;/p&gt;

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

  &lt;div class="agent-session-footer"&gt;
    &lt;span class="agent-session-meta"&gt;
        2 of 2 messages
    &lt;/span&gt;
  &lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;&lt;em&gt;(Link to session: &lt;a href="https://dev.to/agent_sessions/sift-review-offline-hinglish-dm-to-dispatch-engine-opp4cu"&gt;https://dev.to/agent_sessions/sift-review-offline-hinglish-dm-to-dispatch-engine-opp4cu&lt;/a&gt;&lt;br&gt;
)&lt;/em&gt;&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; Sift integrates Google's open-weight &lt;code&gt;gemma2:2b&lt;/code&gt; model locally via Ollama alongside Pydantic structured outputs to extract Hinglish order entities on consumer laptop CPUs with minimal RAM overhead.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Entire:&lt;/strong&gt; Shared the full end-to-end architectural planning, debugging, and local deployment agent session behind Sift using DevRelay / Entire.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of GitHub Copilot:&lt;/strong&gt; Utilized GitHub Copilot inside VS Code to accelerate building the Pydantic validation schema, regex cloud-demo fallback, and SQLite database CRUD operations.&lt;/li&gt;
&lt;/ul&gt;

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