🌾 GramGPT: The AI Panchayat

✨ Inspiration

India’s rural population (65%+) faces daily challenges — limited access to quality education, modern farming knowledge, healthcare awareness, and small-business guidance.

While AI is advancing rapidly, these tools are often designed for urban, English-speaking users with reliable internet. We asked ourselves:
“What if every village had its own AI Panchayat, an offline, multilingual knowledge hub that empowers people with reasoning, not just answers?”

This sparked the creation of GramGPT: a lightweight, fine-tuned local AI agent built on gpt-oss open models, designed for Tier-2/3 towns and rural India.

🏗 What We Built

Four AI Adapters (LoRA fine-tunes)

• 📚 ShikshaGPT – Localized learning assistant for students.
  
• 🌱 KisanGPT – Agricultural copilot for farmers.
  
• 🏥 SwasthyaGPT – Healthcare & wellness guide.
  
• 🛍 VyapaarGPT – Business mentor for small entrepreneurs.
  

Core Features

• Offline Agentic System: Runs fully on-device with vLLM / Ollama, no internet dependency.
  
• Multilingual Support: Fine-tuned on Hindi + regional datasets, with Whisper + Indic-TTS for voice I/O.
  
• Knowledge Augmentation: Integrated with local vector DB (Supabase / SQLite) containing government schemes, weather updates (if online), crop prices, and educational material.
  
• One-stop UI: A clean Next.js app with 4 tabs (Education, Farming, Health, Business), voice-first interface, and chatbot mode.
  

⚙️ How We Built It

• Base Models: gpt-oss-20b for deployment, gpt-oss-120b for few-shot evaluation.
  
• Fine-tuning: Used LoRA adapters with domain datasets (NCERT + AgriStack + Health articles + MSME docs).
  
• Datasets: Mined from open-source (Govt portals, Wikipedia, Indic NLP datasets, WHO health FAQs).
  

🔄 Pipeline

1. Data curation & translation to Hindi/regional languages.
   
2. Fine-tuning with LoRA for each domain.
   
3. Local inference with vLLM/Ollama.
   
4. Voice I/O integrated with Whisper (speech-to-text) + Indic-TTS (text-to-speech).
   
5. UI in Next.js + Tailwind + Supabase.
   

📚 What We Learned

• Reasoning → Information: Fine-tuned models that explain “why” are far more trusted in rural settings.
  
• Multilingual UX matters: Even simple Hinglish + voice output drastically improves adoption.
  
• Local agents shine offline: A 20B model running on a laptop (no internet) is a game-changer for remote areas.
  
• Fine-tuning tradeoffs: Balancing dataset size vs inference speed was key to keeping it lightweight.
  

🚧 Challenges We Faced

• Data Scarcity: Finding clean, domain-specific Hindi datasets was tough; we had to translate and normalize raw sources.
  
• Compute Limits: Training adapters on a laptop wasn’t feasible, we optimized with LoRA + 8-bit quantization.
  
• Voice I/O Latency: Whisper models were slow on CPU; we optimized with faster-whisper + caching.
  
• UX for Rural Users: Designing a UI simple enough for non-tech users (icons, voice, minimal text) required iteration.
  

🔥 In short:
GramGPT bridges the AI divide by giving villages their own reasoning agent — offline, multilingual, and domain-aware.

Built With

• bgeembeddings
• fastapi
• fine-tuning
• framer
• gpt-oss-20b
• lora
• ncert
• next.js
• ollama
• pytorch
• sqlite
• tailwindcss
• whisper