ClinicalGuard AI

An AI-first clinical safety platform that prevents medical errors through multi-agent orchestration, grounded retrieval, and human-in-the-loop escalation.

Architecture

The system uses a cyclic LangGraph graph with five specialised nodes:

1. Intake / Triage — PII/PHI redaction (HIPAA) and deterministic red-flag detection for acute emergencies.
2. Retrieval — ChromaDB vector search with FastMCP JSON-RPC tool calls for drug interactions, allergy cross-reactivity, and lab contraindications.
3. Synthesis — Grounded response generation with explicit chunk citations.
4. Critic / Auditor — Faithfulness and risk scoring. Confidence is computed as C = S_faith × (1 − S_risk).
5. Delivery or Escalation — If confidence ≥ 0.70 the response is cleared; otherwise the system suppresses the AI response and generates an SBAR escalation ticket for physician review.

Results

• 100% overall pipeline accuracy (22/22 benchmark scenarios passing).
• 100% emergency recall rate (5/5 critical crises caught immediately).
• Zero-cost, sovereign architecture — models run locally or on private infrastructure.

Interface

Dual-panel Streamlit application: clinician order entry on the left, live triage and auditor panel with execution traces and confidence breakdowns on the right.