<?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: Srijan </title>
    <description>The latest articles on DEV Community by Srijan  (@srijan07).</description>
    <link>https://dev.to/srijan07</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%2F4176656%2F29718284-f255-4c7b-bca9-7e10ad735049.jpg</url>
      <title>DEV Community: Srijan </title>
      <link>https://dev.to/srijan07</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/srijan07"/>
    <language>en</language>
    <item>
      <title>Organize Cluttered Folders &amp; Summarize Docs with 100% Local Open-Source AI (SLOTH)</title>
      <dc:creator>Srijan </dc:creator>
      <pubDate>Sun, 25 Oct 2026 17:00:03 +0000</pubDate>
      <link>https://dev.to/srijan07/organize-cluttered-folders-summarize-docs-with-100-local-open-source-ai-sloth-443n</link>
      <guid>https://dev.to/srijan07/organize-cluttered-folders-summarize-docs-with-100-local-open-source-ai-sloth-443n</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;
  
  
  🦥 SLOTH — Your Local-First AI Productivity Agent
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;“Your boring tasks. Handled while you rest.”&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;




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

&lt;p&gt;&lt;strong&gt;SLOTH&lt;/strong&gt; is a privacy-first, desktop-friendly web application designed to help developers, students, researchers, and busy knowledge workers conquer digital procrastination. &lt;/p&gt;

&lt;p&gt;Instead of postponing tedious digital chores—like cleaning an overflowing &lt;code&gt;Downloads&lt;/code&gt; folder, figuring out what files named &lt;code&gt;lecture_notes_v2_final_FINAL.pdf&lt;/code&gt; actually contain, or sifting through a 25-page PDF just to find deadlines—SLOTH lets you sit back while a &lt;strong&gt;locally running open-weight AI&lt;/strong&gt; takes care of the boring work.&lt;/p&gt;

&lt;h3&gt;
  
  
  Who gets to be a little lazier because of it?
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Developers &amp;amp; Students:&lt;/strong&gt; Who let hundreds of PDFs, receipts, datasets, and code snippets pile up in unorganized directories.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Researchers &amp;amp; Knowledge Workers:&lt;/strong&gt; Who need instant executive takeaways, action items, and explicitly stated deadlines without reading entire documents end-to-end.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Anyone who wants safe automation:&lt;/strong&gt; SLOTH enforces a strict &lt;strong&gt;two-phase workflow&lt;/strong&gt;—the AI only &lt;em&gt;proposes&lt;/em&gt; a plan. Files are only modified after explicit user review and approval, backed by atomic SQLite journaling and 1-click &lt;strong&gt;conflict-safe Undo&lt;/strong&gt;.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/ZKoe8PbyxkM" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;h3&gt;
  
  
  🎮 Live Walkthrough Highlights
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Smart Folder Organizer:&lt;/strong&gt; Loads a built-in sandbox with 8 messy files (PDF, DOCX, TXT, CSV, JSON, PNG), scans them, and uses local AI to propose human-readable names and subfolders (e.g., &lt;code&gt;Academics/Coursework&lt;/code&gt;, &lt;code&gt;Finance/Receipts&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interactive Plan Table:&lt;/strong&gt; Inline-edit proposed filenames and switch categories via dropdowns before approving changes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reversible Safe Undo:&lt;/strong&gt; 1-click restoration reverts all file operations to their original filenames and paths without data loss.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Document Digest:&lt;/strong&gt; Extracts executive summaries, bulleted key takeaways, checklist action items, and explicitly stated dates/deadlines.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Relax Mode:&lt;/strong&gt; The signature feature featuring a background task queue with laptop-cool concurrency controls and a dynamic &lt;strong&gt;Estimated Time Saved&lt;/strong&gt; counter.&lt;/li&gt;
&lt;/ol&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/srijan-0-7" rel="noopener noreferrer"&gt;
        srijan-0-7
      &lt;/a&gt; / &lt;a href="https://github.com/srijan-0-7/hacktoberfest-dev-3" rel="noopener noreferrer"&gt;
        hacktoberfest-dev-3
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;🦥 SLOTH — Your Local-First AI Productivity Agent&lt;/h1&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;“Your boring tasks. Handled while you rest.”&lt;/strong&gt;&lt;br&gt;
&lt;em&gt;Hacktoberfest 2026 | “Slow and Steady” Track | 100% Local Open-Source AI&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;&lt;a href="https://github.com/srijan-0-7/hacktoberfest-dev-3/LICENSE" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/58dd4da42df86c1864adaeef729174bec2cc4fe94d46a73afd38424fb2da9340/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4c6963656e73652d4d49542d656d6572616c642e737667" alt="License: MIT"&gt;&lt;/a&gt;
&lt;a href="https://python.org" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/7177c3e0c8b7d1fd466fc59b76b13b12388bf2085b45340ad2e213ea52607a98/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f507974686f6e2d332e31322d3264366134662e737667" alt="Python 3.12"&gt;&lt;/a&gt;
&lt;a href="https://fastapi.tiangolo.com" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/7f43b21ad58f961631fee0457061de84145fb52a7a39fe39759e35d8d5dbe23c/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f466173744150492d302e3131312d3430393136632e737667" alt="FastAPI"&gt;&lt;/a&gt;
&lt;a href="https://react.dev" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/9f8032f8bbeca0d317ad49e1ee5cb0188979f9418d510105aa68fc5a0b5a8ea0/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f52656163742d31392e332d3532623738382e737667" alt="React 19"&gt;&lt;/a&gt;
&lt;a href="https://vitejs.dev" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/b1bb86e53cd31c7efa4cfb7fa6d5c9fd31eeed3ca08e7619c799a9241ad4cbb2/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f566974652d382e332d3734633639642e737667" alt="Vite"&gt;&lt;/a&gt;
&lt;a href="https://ollama.com" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/dc5098d4a6742008d8cf08c3d667d92868da60297c5eb9f094859f48973d0b57/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f496e666572656e63652d4f6c6c616d61253230284c6f63616c292d3038316331352e737667" alt="Local Inference"&gt;&lt;/a&gt;
&lt;a href="https://github.com/srijan-0-7/hacktoberfest-dev-3#testing" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/24ad53e7395f27d21fb866de45d9d9d30fb93db9927d0cbc308300741e417ae6/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f54657374732d32322532305061737365642d627269676874677265656e2e737667" alt="Tests Passing"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;📖 Overview&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;SLOTH&lt;/strong&gt; is a privacy-first, desktop-friendly web application designed to help developers, students, researchers, and busy knowledge workers overcome digital procrastination.&lt;/p&gt;
&lt;p&gt;Instead of postponing routine chores—like cleaning a cluttered &lt;code&gt;Downloads&lt;/code&gt; folder, figuring out what messy filenames contain, or reading through 30-page documents to find action items—SLOTH lets you sit back while a &lt;strong&gt;locally running open-weight AI&lt;/strong&gt; does the heavy lifting.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h3 class="heading-element"&gt;🌟 Core Promise&lt;/h3&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Drop the digital clutter. Let your local AI do the boring work. Review once. Relax.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;100% Local Inference:&lt;/strong&gt; Powered by open-weight models via &lt;a href="https://ollama.com" rel="nofollow noopener noreferrer"&gt;Ollama&lt;/a&gt; (&lt;code&gt;qwen2.5:3b&lt;/code&gt;, &lt;code&gt;llama3.2:3b&lt;/code&gt;, &lt;code&gt;mistral:latest&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero Cloud API Fees:&lt;/strong&gt; No OpenAI, Anthropic, or Gemini API keys required for core functionality.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Strict Privacy:&lt;/strong&gt; Your files, filenames, and documents never leave your laptop.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Two-Phase Safety:&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/srijan-0-7/hacktoberfest-dev-3" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The complete project is open-source under the &lt;strong&gt;MIT License&lt;/strong&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Frontend:&lt;/strong&gt; React 19, TypeScript, Vite, Tailwind CSS v4, Lucide Icons.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backend:&lt;/strong&gt; Python 3.12, FastAPI, Uvicorn, Pydantic v2.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Document Processing:&lt;/strong&gt; PyMuPDF, &lt;code&gt;python-docx&lt;/code&gt;, &lt;code&gt;reportlab&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Database:&lt;/strong&gt; SQLite 3 (local durable journal).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Automated Testing:&lt;/strong&gt; 22 passing automated tests in &lt;code&gt;pytest&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;p&gt;SLOTH is built from the ground up around &lt;strong&gt;local inference with open-weight models&lt;/strong&gt; using &lt;strong&gt;Ollama&lt;/strong&gt;:&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Dedicated Local AI Adapter (&lt;code&gt;backend/app/services/ai_service.py&lt;/code&gt;)
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Direct Local Inference:&lt;/strong&gt; Communicates over HTTP with &lt;code&gt;http://127.0.0.1:11434&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dynamic Model Auto-Discovery:&lt;/strong&gt; Queries &lt;code&gt;/api/tags&lt;/code&gt; to discover which models are already pulled on the machine (e.g., &lt;code&gt;qwen2.5:3b&lt;/code&gt;, &lt;code&gt;llama3.2:3b&lt;/code&gt;, &lt;code&gt;mistral:latest&lt;/code&gt;), defaulting gracefully to the active model without hardcoding requirements.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Strict JSON Output:&lt;/strong&gt; Prompts enforce strict Pydantic-validated JSON structures (&lt;code&gt;format="json"&lt;/code&gt;, temperature 0.2).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Truth in AI &amp;amp; Deterministic Fallback:&lt;/strong&gt; If Ollama is offline, SLOTH never fakes AI outputs. Instead, it seamlessly switches to rule-based heuristics and explicitly flags suggestions as &lt;code&gt;[Rule-Based Heuristic (Ollama Offline)]&lt;/code&gt; so the user is always informed.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  2. Multi-Format Text Extractor (&lt;code&gt;backend/app/services/extractor.py&lt;/code&gt;)
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;Safely extracts text and metadata across PDFs, DOCX, TXT, Markdown, CSV, and JSON.&lt;/li&gt;
&lt;li&gt;Detects scanned/image-only PDFs and explicitly flags them for OCR rather than hallucinating content.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  3. Two-Phase Organization &amp;amp; Security Safeguards
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Phase 1 (Preview):&lt;/strong&gt; The AI proposes clean filenames, categories, subcategories, tags, and reasons. The backend sanitizes illegal characters (&lt;code&gt;&amp;lt;&amp;gt;:"/\|?*&lt;/code&gt;) and automatically resolves duplicate target collisions (e.g., &lt;code&gt;Notes (1).pdf&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Phase 2 (Explicit Execution):&lt;/strong&gt; The user approves the batch. The backend re-validates source existence, ensures paths stay strictly inside the target root (rejecting root drives like &lt;code&gt;C:\&lt;/code&gt; and system directories like &lt;code&gt;System32&lt;/code&gt;), moves files atomically, and journals the operation to SQLite.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  4. Conflict-Safe Undo Engine (&lt;code&gt;backend/app/services/undo_service.py&lt;/code&gt;)
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;Every executed operation is logged in the &lt;code&gt;operations&lt;/code&gt; table in SQLite.&lt;/li&gt;
&lt;li&gt;To reverse an action, the engine verifies that the destination file exists and that the original path is currently vacant, preventing accidental overwrites.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  5. Relax Mode Engine (&lt;code&gt;frontend/src/components/RelaxMode.tsx&lt;/code&gt;)
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;Configurable concurrency limiter (default: 2 workers) keeps CPU and RAM lightweight, ensuring laptops stay cool and quiet while processing batches in the background.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;p&gt;This project directly proves why &lt;strong&gt;open innovation and open-weight models&lt;/strong&gt; are superior for everyday personal computing:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Absolute Privacy:&lt;/strong&gt; Your messy personal files, bank receipts, meeting notes, and course syllabi never leave your machine. Closed APIs (like OpenAI or Anthropic) require sending private files over the network to proprietary servers.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero Recurring Costs:&lt;/strong&gt; There are no token charges, credit limits, or unexpected cloud bills. Your hardware does the work offline.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model Sovereignty &amp;amp; Replaceability:&lt;/strong&gt; You are not locked into one vendor. If you want to switch from Qwen to Llama 3.2 or Mistral, you simply change a dropdown in the settings without modifying a single line of application code.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Transparency &amp;amp; Safety:&lt;/strong&gt; Proprietary AI agents often execute black-box operations. SLOTH’s open architecture decouples AI suggestions from deterministic execution—giving the user total review power before a single byte moves on disk.&lt;/li&gt;
&lt;/ol&gt;




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

&lt;p&gt;SLOTH is the ultimate tool for "lazy productivity":&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Folder Cleanup:&lt;/strong&gt; Organizing a cluttered folder of 30 messy downloads manually takes roughly 45 to 60 minutes of opening, reading, renaming, and creating subfolders. With SLOTH, you click &lt;strong&gt;Scan&lt;/strong&gt;, &lt;strong&gt;Analyze with AI&lt;/strong&gt;, review the proposed table in 30 seconds, and hit &lt;strong&gt;Approve&lt;/strong&gt;. &lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Document Reading:&lt;/strong&gt; Instead of reading through a 20-page meeting brief or syllabus, Document Digest isolates&lt;/li&gt;
&lt;/ul&gt;




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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Main Track:&lt;/strong&gt; Hacktoberfest 2026 Open-Source AI Challenge Week 3: &lt;em&gt;Slow and Steady&lt;/em&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>cfbr</title>
      <dc:creator>Srijan </dc:creator>
      <pubDate>Sun, 18 Oct 2026 10:13:12 +0000</pubDate>
      <link>https://dev.to/srijan07/cfbr-dgl</link>
      <guid>https://dev.to/srijan07/cfbr-dgl</guid>
      <description>&lt;div class="ltag__link--embedded"&gt;
  &lt;div class="crayons-story "&gt;
  &lt;a href="https://dev.to/srijan07/from-messy-whatsapp-chits-to-gst-invoices-in-15s-local-ai-with-ollama-qwen-25-p79" class="crayons-story__hidden-navigation-link"&gt;From Messy WhatsApp Chits to GST Invoices in 15s — Local AI with Ollama &amp;amp; Qwen 2.5&lt;/a&gt;


  &lt;div class="crayons-story__body crayons-story__body-full_post"&gt;
      &lt;a href="https://dev.to/srijan07/from-messy-whatsapp-chits-to-gst-invoices-in-15s-local-ai-with-ollama-qwen-25-p79" class="crayons-article__context-note crayons-article__context-note__feed"&gt;&lt;p&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Submission 💸&lt;/p&gt;

&lt;/a&gt;
    &lt;div class="crayons-story__top"&gt;
      &lt;div class="crayons-story__meta"&gt;
        &lt;div class="crayons-story__author-pic"&gt;

          &lt;a href="/srijan07" class="crayons-avatar  crayons-avatar--l  "&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%2Fuser%2Fprofile_image%2F4176656%2F29718284-f255-4c7b-bca9-7e10ad735049.jpg" alt="srijan07 profile" class="crayons-avatar__image" width="460" height="460"&gt;
          &lt;/a&gt;
        &lt;/div&gt;
        &lt;div&gt;
          &lt;div&gt;
            &lt;a href="/srijan07" class="crayons-story__secondary fw-medium m:hidden"&gt;
              Srijan 
            &lt;/a&gt;
            &lt;div class="profile-preview-card relative mb-4 s:mb-0 fw-medium hidden m:inline-block"&gt;
              
                Srijan 
                
                
              
              &lt;div id="story-author-preview-content-4871330" class="profile-preview-card__content crayons-dropdown branded-7 p-4 pt-0"&gt;
                &lt;div class="gap-4 grid"&gt;
                  &lt;div class="-mt-4"&gt;
                    &lt;a href="/srijan07" class="flex"&gt;
                      &lt;span class="crayons-avatar crayons-avatar--xl mr-2 shrink-0"&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%2Fuser%2Fprofile_image%2F4176656%2F29718284-f255-4c7b-bca9-7e10ad735049.jpg" class="crayons-avatar__image" alt="" width="460" height="460"&gt;
                      &lt;/span&gt;
                      &lt;span class="crayons-link crayons-subtitle-2 mt-5"&gt;Srijan &lt;/span&gt;
                    &lt;/a&gt;
                  &lt;/div&gt;
                  &lt;div class="print-hidden"&gt;
                    
                      Follow
                    
                  &lt;/div&gt;
                  &lt;div class="author-preview-metadata-container"&gt;&lt;/div&gt;
                &lt;/div&gt;
              &lt;/div&gt;
            &lt;/div&gt;

          &lt;/div&gt;
          &lt;a href="https://dev.to/srijan07/from-messy-whatsapp-chits-to-gst-invoices-in-15s-local-ai-with-ollama-qwen-25-p79" class="crayons-story__tertiary fs-xs"&gt;&lt;time&gt;Oct 18&lt;/time&gt;&lt;span class="time-ago-indicator-initial-placeholder"&gt;&lt;/span&gt;&lt;/a&gt;
        &lt;/div&gt;
      &lt;/div&gt;

    &lt;/div&gt;

    &lt;div class="crayons-story__indention"&gt;
      &lt;h2 class="crayons-story__title crayons-story__title-full_post"&gt;
        &lt;a href="https://dev.to/srijan07/from-messy-whatsapp-chits-to-gst-invoices-in-15s-local-ai-with-ollama-qwen-25-p79" id="article-link-4871330"&gt;
          From Messy WhatsApp Chits to GST Invoices in 15s — Local AI with Ollama &amp;amp; Qwen 2.5
        &lt;/a&gt;
      &lt;/h2&gt;
        &lt;div class="crayons-story__tags"&gt;
            &lt;a class="crayons-tag  crayons-tag--monochrome " href="/t/devchallenge"&gt;&lt;span class="crayons-tag__prefix"&gt;#&lt;/span&gt;devchallenge&lt;/a&gt;
            &lt;a class="crayons-tag  crayons-tag--monochrome " href="/t/hf26challenge"&gt;&lt;span class="crayons-tag__prefix"&gt;#&lt;/span&gt;hf26challenge&lt;/a&gt;
            &lt;a class="crayons-tag  crayons-tag--monochrome " href="/t/ai"&gt;&lt;span class="crayons-tag__prefix"&gt;#&lt;/span&gt;ai&lt;/a&gt;
            &lt;a class="crayons-tag  crayons-tag--monochrome " href="/t/webdev"&gt;&lt;span class="crayons-tag__prefix"&gt;#&lt;/span&gt;webdev&lt;/a&gt;
        &lt;/div&gt;
      &lt;div class="crayons-story__bottom"&gt;
        &lt;div class="crayons-story__details"&gt;
            &lt;a href="https://dev.to/srijan07/from-messy-whatsapp-chits-to-gst-invoices-in-15s-local-ai-with-ollama-qwen-25-p79#comments" class="crayons-btn crayons-btn--s crayons-btn--ghost crayons-btn--icon-left flex items-center"&gt;
              

              &lt;span class="hidden s:inline"&gt;Add&amp;nbsp;Comment&lt;/span&gt;
            &lt;/a&gt;
        &lt;/div&gt;
        &lt;div class="crayons-story__save"&gt;
          &lt;small class="crayons-story__tertiary fs-xs mr-2"&gt;
            6 min read
          &lt;/small&gt;
        &lt;/div&gt;
      &lt;/div&gt;
    &lt;/div&gt;
  &lt;/div&gt;
&lt;/div&gt;

&lt;/div&gt;


</description>
    </item>
    <item>
      <title>From Messy WhatsApp Chits to GST Invoices in 15s — Local AI with Ollama &amp; Qwen 2.5</title>
      <dc:creator>Srijan </dc:creator>
      <pubDate>Sun, 18 Oct 2026 10:05:17 +0000</pubDate>
      <link>https://dev.to/srijan07/from-messy-whatsapp-chits-to-gst-invoices-in-15s-local-ai-with-ollama-qwen-25-p79</link>
      <guid>https://dev.to/srijan07/from-messy-whatsapp-chits-to-gst-invoices-in-15s-local-ai-with-ollama-qwen-25-p79</guid>
      <description>&lt;h2&gt;
  
  
  BahiPilot: The Local-First AI Back Office for Bharat's Wholesalers
&lt;/h2&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's wholesale markets (&lt;em&gt;mandis&lt;/em&gt;), billions of dollars in daily trade move not through ERP software, but through informal WhatsApp photos of handwritten paper chits (&lt;em&gt;parchis&lt;/em&gt;), voice notes, and loose text messages. &lt;/p&gt;

&lt;p&gt;A small grain distributor or FMCG stockist in a tier-2 trade hub typically spends &lt;strong&gt;2 to 3 hours every evening&lt;/strong&gt; performing tedious, error-prone transcription: deciphering scribbled product abbreviations, checking physical warehouse stock, typing quantities into software, balancing running Khatabook customer ledgers, and printing GST-compliant tax invoices.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;BahiPilot (बही पायलट)&lt;/strong&gt; is an open-source, local-first AI back office built specifically to eliminate this bottleneck for small Indian wholesalers, distributors, and stockists.&lt;/p&gt;

&lt;h3&gt;
  
  
  What BahiPilot Does:
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Multi-Modal Order Ingestion:&lt;/strong&gt; Accepts messy WhatsApp photos, scanned paper chits, digital PDF invoices, or raw Hinglish text messages.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local Open-Weight AI Extraction:&lt;/strong&gt; Uses open-weight models (&lt;code&gt;Qwen 2.5:7b&lt;/code&gt; / &lt;code&gt;Llama 3.2&lt;/code&gt;) via Ollama to parse messy colloquial text into structured JSON line items—with zero marginal token cost and complete data privacy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic Catalogue &amp;amp; Alias Matching:&lt;/strong&gt; Reconciles vernacular terms (&lt;em&gt;"10 bori Basmati Rice 25kg"&lt;/em&gt;, &lt;em&gt;"5 carton Fortune Oil"&lt;/em&gt;, &lt;em&gt;"20 pkt Tata Salt"&lt;/em&gt;) against the business catalogue using token-overlap fuzzy matching and alias indexing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Human-in-the-Loop Review:&lt;/strong&gt; Provides an interactive 4-stage review screen where quantities, unit rates, and customer accounts can be reviewed with ambiguity warning flags before anything is committed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Strict Financial Integrity (Zero LLM Math):&lt;/strong&gt; All subtotals, GST (CGST/SGST/IGST), discount deductions, and totals are computed strictly server-side using Python &lt;code&gt;Decimal&lt;/code&gt; (paise precision). The LLM is never allowed to hallucinate financial figures.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Atomic Ledger &amp;amp; Stock Commitment:&lt;/strong&gt; Zero stock is deducted and zero customer ledger entries are written until the business owner explicitly clicks &lt;strong&gt;Approve&lt;/strong&gt;. Cancellation atomically restores inventory and balances the customer's running ledger.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GST-Ready Printable Invoices &amp;amp; Khatabook:&lt;/strong&gt; Generates clean, printable B2B tax invoices with HSN breakdowns, and maintains a running chronological customer statement with real-time UPI and cash payment recording.&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  Who Would Pay for It?
&lt;/h3&gt;

&lt;p&gt;Small and medium distributors, grain wholesalers (&lt;em&gt;Arhatiyas&lt;/em&gt;), and FMCG stockists handling 30 to 200 orders a day. They operate on thin 2%–5% gross margins and currently either hire data entry clerks (costing ₹15,000–₹25,000/month) or lose hours of their own family time every night balancing paper &lt;em&gt;Bahi Khata&lt;/em&gt; ledgers. A private, local back office that cuts transcription time by 80% pays for itself in days.&lt;/p&gt;




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

&lt;ul&gt;
&lt;li&gt;🎥 &lt;strong&gt;Video Walkthrough:&lt;/strong&gt; Check out our step-by-step video demo showcasing WhatsApp chit ingestion, catalogue reconciliation, atomic approval, and GST invoice printing.
  &lt;iframe src="https://www.youtube.com/embed/XzxOfePZgfc" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/li&gt;
&lt;/ul&gt;




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


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/srijan-0-7" rel="noopener noreferrer"&gt;
        srijan-0-7
      &lt;/a&gt; / &lt;a href="https://github.com/srijan-0-7/hacktoberfest-dev-2" rel="noopener noreferrer"&gt;
        hacktoberfest-dev-2
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;BahiPilot (बही पायलट)&lt;/h1&gt;
&lt;/div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h3 class="heading-element"&gt;Local-First AI Back Office for Small Indian Wholesalers &amp;amp; Distributors&lt;/h3&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;BahiPilot&lt;/strong&gt; streamlines wholesale and distribution operations across Bharat: ingest messy handwritten chits, WhatsApp order screenshots, and PDF invoices; convert them into structured line items using local open-weight AI (Ollama) or Gemini API; reconcile against catalogue and available stock; and execute transactional inventory deduction, customer ledger posting, and GST-ready invoice printing.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;🌟 Key Features&lt;/h2&gt;
&lt;/div&gt;
&lt;ol&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;Dual AI Order Ingestion (Flagship 4-Stage Guided Workflow)&lt;/strong&gt;:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Stage 1 (Input)&lt;/strong&gt;: Upload invoices/receipts (PNG, JPEG, PDF) or paste informal order text (Hinglish/English chits, e.g., &lt;em&gt;"10 bori Basmati Rice 25kg, 5 carton Fortune Oil 1L, 20 packet Tata Salt"&lt;/em&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stage 2 (Extraction)&lt;/strong&gt;: Extracts text via PyMuPDF or Tesseract OCR, prompts local LLMs (Ollama &lt;code&gt;qwen2.5:7b&lt;/code&gt; / &lt;code&gt;llama3.2&lt;/code&gt;) or Google Gemini (&lt;code&gt;gemini-2.5-flash&lt;/code&gt;), strictly adhering to structured JSON schemas.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stage 3 (Review &amp;amp; Reconcile)&lt;/strong&gt;: Interactive review table matching extracted aliases…&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ol&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/srijan-0-7/hacktoberfest-dev-2" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;&lt;strong&gt;Repository:&lt;/strong&gt; &lt;a href="https://github.com/srijan-0-7/hacktoberfest-dev-2" rel="noopener noreferrer"&gt;https://github.com/srijan-0-7/hacktoberfest-dev-2&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Architecture Highlights:
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Frontend:&lt;/strong&gt; React 19, TypeScript, Vite, Tailwind CSS (custom Midnight Navy &amp;amp; Emerald Green theme), TanStack Query v5, Recharts, Lucide Icons.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backend:&lt;/strong&gt; Python 3.11+, FastAPI, Pydantic v2, SQLAlchemy 2.0, SQLite (with WAL mode &amp;amp; strict &lt;code&gt;PRAGMA foreign_keys = ON&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AI Engine:&lt;/strong&gt; Pluggable provider architecture:

&lt;ul&gt;
&lt;li&gt;Local primary: Ollama REST provider (&lt;code&gt;qwen2.5:7b&lt;/code&gt; / &lt;code&gt;llama3.2&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;Cloud fallback: Google Gemini API (&lt;code&gt;gemini-3.8-flash&lt;/code&gt;)&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Document Processing:&lt;/strong&gt; PyMuPDF (&lt;code&gt;fitz&lt;/code&gt;) for native PDF stream extraction, Pillow for preprocessing, and optional Tesseract OCR.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Production Ready:&lt;/strong&gt; Pre-configured &lt;code&gt;Dockerfile&lt;/code&gt; (multi-stage build) and &lt;code&gt;render.yaml&lt;/code&gt; for 1-click cloud deployment.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;h3&gt;
  
  
  1. Dual AI Provider with Structured Pydantic Schemas
&lt;/h3&gt;

&lt;p&gt;BahiPilot implements an extensible &lt;code&gt;BaseAIProvider&lt;/code&gt; interface. The prompt strictly limits the LLM's task to linguistic extraction, injecting the business's real catalogue and customer list as context:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;class&lt;/span&gt; &lt;span class="nc"&gt;ExtractionRawLLMOutput&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;BaseModel&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;customer_name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;Optional&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nb"&gt;str&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="bp"&gt;None&lt;/span&gt;
    &lt;span class="n"&gt;items&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;List&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;RawItemExtraction&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Field&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;default_factory&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nb"&gt;list&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;notes&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;List&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nb"&gt;str&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Field&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;default_factory&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nb"&gt;list&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;overall_needs_review&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;bool&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;By requesting structured JSON output and validating via Pydantic schemas, we eliminate hallucinated formats while handling diverse vernacular inputs (Hinglish, mixed English/Hindi names, informal units).&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Deterministic Alias &amp;amp; Catalogue Matcher
&lt;/h3&gt;

&lt;p&gt;In Indian trade, the same product is called ten different names: &lt;em&gt;"Chawal"&lt;/em&gt;, &lt;em&gt;"Basmati 25kg"&lt;/em&gt;, &lt;em&gt;"Rice bag"&lt;/em&gt;, or &lt;em&gt;"India Gate Classic"&lt;/em&gt;. &lt;/p&gt;

&lt;p&gt;Instead of trusting an LLM to guess database foreign keys, BahiPilot uses a tiered deterministic matching algorithm:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Exact SKU and Product Name normalization&lt;/li&gt;
&lt;li&gt;Exact Alias array lookup&lt;/li&gt;
&lt;li&gt;Token coverage &amp;amp; overlap ratios:
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="c1"&gt;# Query coverage: what percentage of the requested item's words exist in this product?
&lt;/span&gt;&lt;span class="n"&gt;overlap&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;query_tokens&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;intersection&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;prod_tokens&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;query_coverage&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;overlap&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;query_tokens&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;jaccard&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;overlap&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;query_tokens&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;union&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;prod_tokens&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
&lt;span class="n"&gt;score&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;max&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;query_coverage&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mf"&gt;0.82&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;jaccard&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Unmatched or low-confidence items are surfaced with visual warning badges on the review screen, forcing human-in-the-loop resolution before committing to stock.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Strict Server-Side Financial Math &amp;amp; Transactional State
&lt;/h3&gt;

&lt;p&gt;To guarantee financial trust:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Zero arithmetic is performed by the LLM. Line subtotals, CGST/SGST splits, and grand totals are calculated with Python &lt;code&gt;Decimal&lt;/code&gt; in paise accuracy.&lt;/li&gt;
&lt;li&gt;Orders start in &lt;code&gt;DRAFT&lt;/code&gt; or &lt;code&gt;PENDING_REVIEW&lt;/code&gt;. Stock deduction and ledger postings execute &lt;strong&gt;only&lt;/strong&gt; upon explicit user &lt;code&gt;APPROVE&lt;/code&gt; action in an atomic database transaction. If cancelled, an automated stock restoration and ledger debit reversal is logged.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;p&gt;Open-source and open-weight AI are not just philosophical preferences for Bharat's MSMEs—&lt;strong&gt;they are economic and operational necessities&lt;/strong&gt;:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Zero Marginal Inference Cost:&lt;/strong&gt;&lt;br&gt;&lt;br&gt;
Proprietary cloud APIs charging per token or per image are economically unfeasible for small businesses operating on 2%–4% margins. A distributor processing 150 chits a day would face unpredictable recurring monthly SaaS bills. With open-weight models like Qwen 2.5 running on a basic local shop desktop via Ollama, &lt;strong&gt;marginal inference cost is exactly ₹0.00&lt;/strong&gt;.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Absolute Data Sovereignty &amp;amp; Confidentiality:&lt;/strong&gt;&lt;br&gt;&lt;br&gt;
Indian wholesalers guard their customer directories, negotiated volume discounts, and turnover numbers as core trade secrets. Small business owners are rightfully wary of transmitting unredacted order photos and customer lists to third-party cloud servers. Local open weights keep 100% of the data inside the four walls of the shop.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Offline Continuity in Trade Mandis:&lt;/strong&gt;&lt;br&gt;&lt;br&gt;
Wholesale grain and spice markets often suffer from spotty mobile networks and intermittent broadband. Because BahiPilot's inference and database run entirely on localhost, the distributor can continue scanning orders, adjusting stock, and issuing invoices even when the internet is completely down.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Open Vocabulary &amp;amp; Model Independence:&lt;/strong&gt;&lt;br&gt;&lt;br&gt;
Open weights can be fine-tuned or prompt-engineered for regional Indian languages (Hindi, Marathi, Gujarati, Tamil) and unique market jargon without vendor lock-in or fear of arbitrary API deprecation.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;




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

&lt;p&gt;This entire project was architected, implemented, and debugged through an end-to-end AI pair programming session using Antigravity. &lt;/p&gt;

&lt;p&gt;Key milestones accomplished during the agent session:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;System Architecture &amp;amp; Domain Modeling:&lt;/strong&gt; Engineered the SQLite multi-table schema (&lt;code&gt;customers&lt;/code&gt;, &lt;code&gt;products&lt;/code&gt;, &lt;code&gt;orders&lt;/code&gt;, &lt;code&gt;order_items&lt;/code&gt;, &lt;code&gt;ledger_entries&lt;/code&gt;) with strict foreign keys and transactional consistency.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dual AI Provider Integration:&lt;/strong&gt; Built the pluggable &lt;code&gt;BaseAIProvider&lt;/code&gt; abstraction supporting local Ollama (&lt;code&gt;qwen2.5:7b&lt;/code&gt; / &lt;code&gt;llama3.2&lt;/code&gt;) and Gemini cloud fallback with structured JSON validation via Pydantic v2.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic Token-Overlap Matcher:&lt;/strong&gt; Developed the fuzzy alias matching engine to reconcile colloquial Indian trade terms (&lt;em&gt;"10 bori Basmati Rice"&lt;/em&gt;, &lt;em&gt;"5 carton Fortune Oil"&lt;/em&gt;) with catalogue items at 85%–95% confidence without calling external APIs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Full-Stack Polish &amp;amp; Verification:&lt;/strong&gt; Created the responsive React 19 + Tailwind CSS frontend, wired TanStack Query endpoints, resolved Vite URL routing, and verified atomic approval, inventory deduction, and GST invoice generation.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  My First User
&lt;/h2&gt;

&lt;p&gt;To ground BahiPilot in reality, I tested the order import pipeline with &lt;strong&gt;Ramesh Gupta&lt;/strong&gt;, who operates a wholesale grocery and provisions store (&lt;em&gt;Gupta Kirana &amp;amp; Provisions&lt;/em&gt;) supplying 45 retail kirana shops in Maharashtra.&lt;/p&gt;

&lt;h3&gt;
  
  
  What He Tested:
&lt;/h3&gt;

&lt;p&gt;Ramesh pasted a typical evening WhatsApp order list sent by one of his regular buyers:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"10 bori Basmati Rice 25kg, 5 carton Fortune Oil 1L, 20 pkt Tata Salt, 15 packet Maggi. Urgent delivery by 4 PM."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h3&gt;
  
  
  What Happened:
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;BahiPilot automatically recognized the buyer (&lt;em&gt;Gupta Kirana General Stores&lt;/em&gt;) and matched all 4 products against his catalogue inventory with 85%–95% confidence.&lt;/li&gt;
&lt;li&gt;The system flagged that his catalogue unit for Maggi was a 24-pack box, allowing him to verify the pack size with one click.&lt;/li&gt;
&lt;li&gt;Upon clicking &lt;strong&gt;Approve&lt;/strong&gt;, the stock deducted from his digital godown, the invoice was posted to the buyer's running ledger, and a GST-compliant print bill was ready in under 15 seconds.&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  His Feedback:
&lt;/h3&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"Every evening after 8 PM, I sit with a notebook and spend an hour and a half making bills and checking who owes how much. If this software can take my WhatsApp orders and prepare my bills and Khatabook without needing an internet connection, it saves me at least 10 hours a week. I don't want my customer rates going to foreign clouds, so keeping it on my shop laptop is a huge relief."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;




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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/strong&gt; (Main Category)&lt;/li&gt;
&lt;/ul&gt;




&lt;h3&gt;
  
  
  Tech Stack Summary
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Frontend:&lt;/strong&gt; React 19, TypeScript, Tailwind CSS, Vite, TanStack Query&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backend:&lt;/strong&gt; FastAPI, Pydantic v2, SQLAlchemy, SQLite&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AI Models:&lt;/strong&gt; Ollama (&lt;code&gt;qwen2.5:7b&lt;/code&gt;), Google Gemini API (&lt;code&gt;gemini-3.8-flash&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Document Processing:&lt;/strong&gt; PyMuPDF, Pillow, Tesseract OCR&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deployment:&lt;/strong&gt; Docker, Render Blueprint (&lt;code&gt;render.yaml&lt;/code&gt;), Localtunnel&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;Built with ❤️ for Indian MSMEs and the open-source AI community.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>webdev</category>
    </item>
    <item>
      <title>FieldLens 🌿: Exploring the Outdoors with Gemma-Powered Vision AI</title>
      <dc:creator>Srijan </dc:creator>
      <pubDate>Sun, 11 Oct 2026 16:15:20 +0000</pubDate>
      <link>https://dev.to/srijan07/fieldlens-exploring-the-outdoors-with-gemma-powered-vision-ai-5f35</link>
      <guid>https://dev.to/srijan07/fieldlens-exploring-the-outdoors-with-gemma-powered-vision-ai-5f35</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;
  
  
  FieldLens 🌿 — Gemma-Powered Nature Identifier
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;“Step outside. Discover what's around you.”&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

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

&lt;p&gt;&lt;strong&gt;FieldLens&lt;/strong&gt; is a lightweight, privacy-preserving nature observation web application designed to encourage people to close their laptops, step outside, and engage with the living world around them. &lt;/p&gt;

&lt;p&gt;Whether exploring a local trail, walking through a neighborhood park, or examining flora and fauna in your backyard, FieldLens lets you photograph any plant, flower, tree, bird, mushroom, or insect and receive grounded, cautious insights powered by &lt;strong&gt;Google Gemma&lt;/strong&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  Key Features
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Mobile-First Field Capture:&lt;/strong&gt; Take photos directly through your smartphone camera (&lt;code&gt;capture="environment"&lt;/code&gt;) or upload existing images (JPEG, PNG, WebP) with automatic client-side compression.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Grounded, Cautious Identification:&lt;/strong&gt; Unlike generic chatbots that hallucinate species names, FieldLens is instructed to distinguish visible morphological evidence from assumptions. When diagnostic traits are ambiguous or missing, it identifies the broadest defensible category (&lt;em&gt;flowering plant&lt;/em&gt;, &lt;em&gt;bird&lt;/em&gt;, &lt;em&gt;tree&lt;/em&gt;, &lt;em&gt;mushroom&lt;/em&gt;, &lt;em&gt;insect&lt;/em&gt;, or &lt;em&gt;unknown organism&lt;/em&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Calibrated Uncertainty Ratings:&lt;/strong&gt; Honest qualitative uncertainty levels (&lt;strong&gt;Low&lt;/strong&gt;, &lt;strong&gt;Moderate&lt;/strong&gt;, &lt;strong&gt;High&lt;/strong&gt;) with clear explanations—no fabricated percentage scores.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Actionable Outdoor Guidance:&lt;/strong&gt; Practical outdoor observation tips (e.g. leaf arrangement, bark texture, scent check) and safety warnings.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Offline Personal Nature Journal:&lt;/strong&gt; Save sightings directly to your browser's local memory (&lt;code&gt;localStorage&lt;/code&gt; for metadata, &lt;code&gt;IndexedDB&lt;/code&gt; for high-resolution images). Persists across page refreshes with zero tracking and zero login requirements.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero Server Retention:&lt;/strong&gt; Uploaded images are processed in-memory and discarded immediately upon analysis.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;p&gt;Watch FieldLens in action analyzing specimens and saving observations in real time:&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/YIhn4hLXFb0" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Direct Video Link:&lt;/strong&gt; &lt;a href="https://youtu.be/YIhn4hLXFb0?si=0_rFT6y7QsZZdGr9" rel="noopener noreferrer"&gt;https://youtu.be/YIhn4hLXFb0?si=0_rFT6y7QsZZdGr9&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Walkthrough &amp;amp; User Journey
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Landing &amp;amp; Ethics:&lt;/strong&gt; Introduces &lt;em&gt;Leave No Trace&lt;/em&gt; outdoor ethics and transparency disclosures.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Capture &amp;amp; Analyze:&lt;/strong&gt; Select or photograph a specimen; Google Gemma Vision evaluates leaf structure, plumage, or growth habit in real time.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Structured Card:&lt;/strong&gt; Renders common name, scientific name (if warranted), uncertainty rating, visual evidence, outdoor tips, and cautions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Field Journal:&lt;/strong&gt; Browse, search, filter, and inspect your saved observations over time.&lt;/li&gt;
&lt;/ol&gt;




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

&lt;p&gt;The complete source code is public and open-source:&lt;/p&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/srijan-0-7" rel="noopener noreferrer"&gt;
        srijan-0-7
      &lt;/a&gt; / &lt;a href="https://github.com/srijan-0-7/hacktoberfest-dev-1" rel="noopener noreferrer"&gt;
        hacktoberfest-dev-1
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;FieldLens 🌿 — Gemma-Powered Nature Identifier&lt;/h1&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;“Step outside. Discover what's around you.”&lt;/strong&gt;&lt;br&gt;
&lt;em&gt;Built for Hacktoberfest 2026 Week 1 DEV Community Challenge: “Touch Grass”&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;FieldLens is a lightweight, privacy-preserving nature observation web application that uses Google's open-weights &lt;strong&gt;Gemma&lt;/strong&gt; vision-language models to help outdoor explorers, hikers, and curious backyard observers identify and learn about plants, flowers, birds, trees, mushrooms, and insects in their environment.&lt;/p&gt;
&lt;p&gt;Unlike generic image classifiers or ungrounded chatbot wrappers, FieldLens is designed around a &lt;strong&gt;cautious observation philosophy&lt;/strong&gt;: it examines visible morphological traits, distinguishes evidence from assumptions, qualifies its uncertainty honestly, provides practical outdoor tips, and reminds users never to consume or disturb unknown wildlife.&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;📺 Video Demo&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;Watch FieldLens in action identifying specimens and saving observations:&lt;/p&gt;
&lt;p&gt;&lt;a href="https://youtu.be/YIhn4hLXFb0?si=0_rFT6y7QsZZdGr9" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/569b938d28e7c5aaf8cbd635b5833322162c1ada39e042b00291a8327ce6b938/68747470733a2f2f696d672e796f75747562652e636f6d2f76692f5949686e34684c584662302f302e6a7067" alt="FieldLens Video Demo"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;👉 &lt;strong&gt;&lt;a href="https://youtu.be/YIhn4hLXFb0?si=0_rFT6y7QsZZdGr9" rel="nofollow noopener noreferrer"&gt;Watch on YouTube&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;🍃 Connection to “Touch Grass”&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;The theme of Hacktoberfest Week 1 is &lt;strong&gt;“Touch Grass”&lt;/strong&gt;: leveraging open-source AI to encourage people to disconnect from repetitive digital routines and…&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/srijan-0-7/hacktoberfest-dev-1" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;&lt;strong&gt;Repository:&lt;/strong&gt; &lt;a href="https://github.com/srijan-0-7/hacktoberfest-dev-1" rel="noopener noreferrer"&gt;https://github.com/srijan-0-7/hacktoberfest-dev-1&lt;/a&gt;&lt;/p&gt;




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

&lt;h3&gt;
  
  
  1. Open-Source AI Model: Google Gemma Vision
&lt;/h3&gt;

&lt;p&gt;FieldLens is powered by &lt;strong&gt;Google Gemma 4 Vision&lt;/strong&gt; (&lt;code&gt;models/gemma-4-26b-a4b-it&lt;/code&gt;), an open-weights multimodal vision-language model developed by Google DeepMind.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;System Prompt Design:&lt;/strong&gt; The model is bound by a strict persona prompt as &lt;em&gt;FieldLens&lt;/em&gt;, a cautious nature assistant that refuses to invent Latin binomials or claim species certainty unless distinguishing morphological traits are visibly present in the photograph.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Resilient Multi-Tier JSON Parsing:&lt;/strong&gt; To handle reasoning traces and varying response patterns, the parser dynamically extracts valid schemas from direct JSON, markdown-fenced code blocks (&lt;code&gt;&lt;/code&gt;`&lt;code&gt;json&lt;/code&gt;), and structured key-value bullet points.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  2. Frontend &amp;amp; Local Persistence
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;React 18 + TypeScript + Vite:&lt;/strong&gt; Clean, responsive, and mobile-friendly interface styled with Tailwind CSS and Lucide icons.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hybrid Browser Storage:&lt;/strong&gt; Solved browser storage quota issues by using &lt;code&gt;localStorage&lt;/code&gt; for structured observation metadata and &lt;code&gt;IndexedDB&lt;/code&gt; (via &lt;code&gt;idb-keyval&lt;/code&gt;) for storing full-resolution photograph blobs locally.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  3. Serverless API Architecture
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Serverless API Handler (&lt;code&gt;api/identify.ts&lt;/code&gt;):&lt;/strong&gt; Lightweight Node.js endpoint with built-in retry logic and timeout safeguards, ensuring API keys stay strictly server-side.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Integrated Vite Dev Middleware:&lt;/strong&gt; The local Vite development server automatically intercepts &lt;code&gt;/api/identify&lt;/code&gt; using &lt;code&gt;dotenv&lt;/code&gt;, offering a unified developer experience with single-command startup (&lt;code&gt;npm run dev&lt;/code&gt;).&lt;/li&gt;
&lt;/ul&gt;




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

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Scientific Grounding over Marketing Hype:&lt;/strong&gt; Closed commercial APIs often push conversational models to appear authoritative, leading to dangerous hallucinations in the natural world (such as incorrectly declaring a wild fungus edible). Open-weights models like Gemma allow us to carefully craft system boundaries, inspect behavioral characteristics, and build tools centered on scientific caution.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Democratizing Outdoor Literacy:&lt;/strong&gt; Nature belongs to everyone. Open innovation makes it possible to build free, privacy-first field companions without locking basic biological education behind monthly paywalls, tracking cookies, or mandatory user accounts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Auditability &amp;amp; Longevity:&lt;/strong&gt; An open-weights ecosystem guarantees that research tools and conservation applications can be audited, adapted, and run independently without reliance on a single proprietary vendor.&lt;/li&gt;
&lt;/ol&gt;




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

&lt;p&gt;FieldLens was architected and built in collaboration with &lt;strong&gt;Antigravity&lt;/strong&gt;, an autonomous agentic pair-programmer powered by Gemini:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;End-to-end multimodal API testing and prompt calibration with real botanical specimens.&lt;/li&gt;
&lt;li&gt;TypeScript schema design and multi-tier extractor implementation.&lt;/li&gt;
&lt;li&gt;Automated testing suite (&lt;code&gt;scripts/test-api.mjs&lt;/code&gt;) verifying inference and response validity.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Touch Grass Challenge (Week 1)&lt;/strong&gt;: Encouraging real-world outdoor engagement, nature observation, and Leave No Trace ethics through open-source AI.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Google Gemma Challenge&lt;/strong&gt;: Practical application of Google Gemma open-weights multimodal vision models to solve a real-world environmental education challenge.&lt;/li&gt;
&lt;/ul&gt;

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