ATLAS — Healthcare AI Safety Infrastructure

ATLAS is an experimental healthcare-AI safety and interoperability stack built around a simple principle: clinical AI needs auditable infrastructure around the model, not just the model itself.

What I built

ATLAS combines:

• Consent-first access with scope, purpose, expiration, revocation, and emergency break-glass handling
• SHA-256 hash-chained audit events with validation and tamper detection
• SMART on FHIR identity and OAuth2/PKCE integration
• FHIR R4 support for Patient, Observation, Condition, MedicationRequest, Encounter, and Bundle resources
• MCP servers for audit and consent
• A2A-style typed agent handoffs with SHARP context propagation
• A deterministic rule-based triage fallback when an LLM is unavailable
• A REST API and interactive local dashboard

Verification

The repository currently documents 39 tests across six suites covering audit integrity, consent, triage, A2A handoffs, SHARP propagation, AI fallback behavior, and core safety primitives.

The demo includes an end-to-end emergency-triage scenario and reports the audit-chain validation result.

Important scope

ATLAS is a research/reference implementation and local evaluation system, not a production clinical deployment. It does not provide medical advice or replace qualified clinical judgment.

Source

GitHub: https://github.com/esskslifetech/ATLAS-Healthcare-AI-Safety-Infrastructure

Built With

• a2a
• fhir-r4
• jest
• mcp
• node.js
• oauth2
• sha-256
• smart-on-fhir
• typescript