A Retrieval-Augmented Generation (RAG) system that helps everyday Indians navigate the maze of government welfare schemes — in Hinglish, over WhatsApp.
The Problem
India has 200+ central government schemes (PM Kisan, Ayushman Bharat, PM Awas Yojana, etc.) but most citizens have no idea which ones they qualify for or how to apply. Existing portals are English-only, bureaucratically worded, and have terrible UX.
The goal: make this information accessible to people who speak Hinglish and use WhatsApp daily.
Architecture
User (WhatsApp) → Twilio → FastAPI Backend
↓
Hinglish Query Preprocessing
↓
Pinecone Vector Search
↓
Scheme Knowledge Base (200+ schemes)
↓
LLM Response Generation
↓
Hinglish Response → User
What I Built
- Vector retrieval pipeline using Pinecone for semantic search over 200+ scheme documents, chunked and embedded
- FastAPI backend handling Twilio webhook events, query preprocessing, and response formatting
- Hinglish-aware prompting — the LLM responds in the language the user writes in
- Docker + Render deployment for zero-downtime production serving
- Twilio WhatsApp integration — no app install required, runs on a platform users already have
Impact
- Designed to scale to 1,000+ daily users
- Covers schemes across agriculture, healthcare, housing, education, and employment
- Response latency: under 3 seconds end-to-end