Inspiration

PMJAY agents spend 80% of call time manually searching policy documents Beneficiaries face long wait times and inconsistent answers No existing tool gives agents real-time, document-grounded assistance during live calls

What it does

Transcribes customer speech using Whisper in real time Classifies intent, urgency, and sentiment simultaneously using fine-tuned DistilBERT Retrieves accurate answers from official PMJAY documents using RAG Auto-resolves routine queries without human intervention Generates AI action plans for complaints and escalations using Llama 3.3 Reads answers aloud to the agent via text-to-speech

How we built it

Whisper (tiny, int8) for speech transcription Three parallel DistilBERT classifiers for intent, urgency, sentiment ChromaDB + BM25 hybrid search over official PMJAY PDFs FastAPI backend with SSE streaming and SQLite React dashboard with real-time WebSocket updates Groq API for action plan generation

Challenges we ran into

Getting parallel BERT inference to work reliably on CPU RAG returning irrelevant answers for off-topic queries Browser audio encoding causing Whisper decode errors Keeping the entire pipeline under 10 seconds on CPU hardware

Accomplishments that we're proud of

Full end-to-end pipeline working in under 10 seconds Auto-resolving ~70% of routine queries with zero human intervention Every answer grounded in official NHA documents — no hallucination AI action plans specific to each case using live RAG context

What we learned

RAG quality depends entirely on document quality and chunking strategy Parallel inference significantly reduces latency even on CPU Autonomous resolution requires careful intent-based routing, not just confidence scores Groq is remarkably fast for real-time LLM inference in production

What's next for ClaimDesk AI - Automating Claim Insights for Agents

Real phone call integration via Twilio SIP Hindi and regional language support using Whisper multilingual Live hospital empanelment lookup from NHA database Conversation memory for multi-turn interactions Expanded knowledge base with state-specific PMJAY guidelines

Built With

  • bm25
  • chromadb
  • fastapi
  • groq-api-(llama-3.3)
  • huggingface-transformers-(distilbert)
  • openai-whisper
  • python
  • react
  • server-sent-events
  • sqlite
  • web-speech-api
  • websocket
Share this project:

Updates