CureWiseAgentic RAG for Healthcare
A production-grade healthcare platform that combines fine-tuned vision models, agentic RAG, and LLMs to deliver medical analytics, report parsing, and conversational AI, all dockerized and deployed end to end.
Headline results
- Surface
- Deployed platform
- Capabilities
- Chat · RAG · vision · DB agents
- Retrieval
- Pinecone + LlamaParse
- Stack
- FastAPI · React · Docker
System architecture
Problem
Doctors and patients needed one interface to parse medical reports, diagnose skin conditions from images, and book appointments, without the hallucinations that kill clinical trust.
Approach
Built an agentic RAG architecture: LangGraph routes each request to the right tool: a fine-tuned vision model for image diagnosis, a Pinecone-backed retriever grounded in PostgreSQL patient history, or a calendar agent. Every LLM answer is constrained to retrieved context and dockerized for reproducible deployment.
Impact
Zero-hallucination responses grounded in real-time PostgreSQL patient history. Fine-tuned vision models diagnose acne and rashes from uploads. Agents hit a calendar API to book appointments directly from chat.
Decisions & tradeoffs
Agentic routing over one mega-prompt
Diagnosis, retrieval, and booking have different tools and different failure modes. LangGraph routes each request to a specialist path instead of stretching a single prompt across all three jobs.
Every answer constrained to retrieved context
Clinical trust dies with the first hallucination. Responses are grounded in records fetched live from PostgreSQL; when retrieval comes back empty, the system says so instead of guessing.
Docker from day one
Healthcare deployments demand reproducibility. The entire stack ships as one command, so the demo environment and production behave identically.
Build spec
- Status
- Production-grade · open source
- Orchestration
- LangGraph agentic router · 3 tool paths
- Data layer
- Pinecone vectors + live PostgreSQL records
- Models
- Fine-tuned vision classifier + GPT-3.5
- Deployment
- Dockerized · one-command reproducible
System notes
- Agentic routing with LangGraph across diagnosis, retrieval, and booking tools
- Vision models fine-tuned to classify skin conditions from user uploads
- RAG grounded in live PostgreSQL patient records, no hallucinated history
- Fully dockerized for one-command, reproducible deployment
What this does not show
- Breadth is the point and the weakness: it is a wide product rather than one sharp measured claim.
Stack
LangChain · LangGraph · Pinecone · Docker · PostgreSQL · GPT-3.5