Multi-Agent AI Hospital Simulation

Async FastAPI backend with JWT auth, LangGraph state machine with conditional routing across GP, specialist, and helper agents, real-time SSE streaming to React frontend; serves live production traffic on a 4GB DigitalOcean droplet.

Key Highlights

• Engineered async FastAPI backend with JWT auth, LangGraph state machine with conditional routing across GP, specialist, and helper agents, and real-time SSE streaming to React frontend; system serves live production traffic on a 4GB DigitalOcean droplet.
• Designed Isolated Message Context Architecture (IMCA) with 9 specialty-partitioned ChromaDB vector stores (BAAI/bge-large-en-v1.5) and bridge arrays for inter-agent context passing, reducing per-consultation token consumption by 81%.
• Resolved production context-corruption: identified that LangGraph's MemorySaver() is in-process and non-shared across Gunicorn workers; enforced single-worker constraint via systemd service config, eliminating cross-request state loss.