Governed Healthcare Agent

Multi-agent LLM system for clinical decision support with guardrails, audit logging, and retrieval-augmented generation over medical knowledge bases.

A governed multi-agent system for clinical decision support. Built on LangGraph for agent orchestration, with pgvector-backed retrieval over medical literature, and a policy layer that enforces guardrails on what the system can and cannot recommend.

Motivation

Clinical LLM deployments fail in two ways: hallucination on medical facts, and uncontrolled scope (the model does things it shouldn’t). This project treats both as engineering problems, not prompting problems.

Architecture

User query
  → Intent classifier (LangGraph node)
  → Retrieval agent (pgvector similarity search)
  → Synthesis agent (GPT-4 with retrieved context)
  → Policy gate (rule-based + classifier)
  → Response
  → Audit log (PostgreSQL)

The policy gate runs after synthesis and before delivery. It checks the response against a set of constraints — no specific dosage recommendations, no diagnoses — and either passes, flags for review, or blocks.

Status

Retrieval and synthesis pipeline complete. Policy gate in active development. Evaluating RAGAS metrics for retrieval quality baseline.