20241st place
Clinical AI Chatbot
Winner at Afya Hacks: a HIPAA-minded clinical agent, built with a team of three, that helps doctors with patient management and diagnosis - LangChain over a FastAPI backend, Dockerized with CI/CD on Cloud Run.
Overview
Built for Afya Hacks with a team of three - and took first place - this is a clinical assistant for doctors, helping with patient management, symptom assessment, and routing cases to the right level of care.
The defining constraint was privacy. Instead of calling a hosted LLM API, the agent runs entirely on local models served through Ollama, so no patient-adjacent text ever leaves the infrastructure - the pragmatic answer to HIPAA concerns in a healthcare setting. On the evaluation set, the local pipeline reached 88% accuracy.
A FastAPI backend orchestrates the LangChain flows over a PostgreSQL store. The whole system is containerized and ships through a CI/CD pipeline to Google Cloud Run, so a merge to main becomes a deployed revision without manual steps.
Highlights
- First place at Afya Hacks, built by a team of three.
- Doctor-facing: patient management, symptom assessment, and case routing.
- Local-first inference via Ollama - no PHI-adjacent data sent to third-party model APIs.
- LangChain orchestration on a FastAPI backend with a PostgreSQL store.
- 88% accuracy on the evaluation set using local models.
- Dockerized, with CI/CD deploying to Google Cloud Run.