Solutions / Healthcare

Prove every
clinical AI decision

AI agents in healthcare access PHI, recommend treatments, and make clinical decisions. Project AIR provides the cryptographic proof that every access was authorized, every decision was logged, and every action can be audited.

Get Started Contact Sales

Capabilities

PHI Access Audit Trail

Every patient record access by an AI agent is captured as a signed intent capsule with BLAKE3 chain hashing and Ed25519 signatures. No PHI touches go unrecorded.

45 CFR 164.312(b) Compliance

Audit controls required by the HIPAA Security Rule are satisfied through tamper-evident forensic chains. Each record is cryptographically linked to the previous one.

Integrity Controls

45 CFR 164.312(c) integrity requirements are met by the signed chain architecture. Any modification to any record is detectable at the byte level.

Breach Detection

16 OWASP-mapped detectors run locally with zero data exfiltration. Prompt injection, data leakage, privilege escalation, and tool misuse are flagged in real time.

Human-in-the-Loop

Auth0-verified containment (Layer 3) ensures sensitive clinical actions require human approval via JWT-verified step-up authentication. Deny rules override step-up rules.

Compliance Reporting

Generate HIPAA evidence reports, EU AI Act Article 72 reports, and NIST AI RMF assessments directly from your signed forensic chain.

HL7v2 Clinical Evidence

Every ADT, ORM, ORU, MDM message your clinical AI agent handles is parsed and recorded as a signed capsule. PHI is redacted by default to minimize exposure; BAA required for all clinical deployments.

FHIR R4 Structured Evidence

HL7v2 segments are mapped to FHIR R4 resources (Patient, Observation, ServiceRequest, DiagnosticReport) using the HL7-published spec models. Auditors see structured clinical data with proper coding system attribution.

Quick Start

Instrument a clinical agent in 4 lines

from airsdk import AIRRecorder

recorder = AIRRecorder("clinical-chain.jsonl")

recorder.tool_start(tool_name="ehr_read", tool_args={"patient_id": "P-4821"})
recorder.tool_end(tool_output="HbA1c: 7.2, BP: 138/88")

recorder.llm_start(prompt="Review labs and recommend next steps")
recorder.llm_end(response="Recommend titrating metformin...")

recorder.agent_finish(final_output="Clinical review complete")

Every call generates a signed, tamper-evident record. Run air demo --healthcare to see it in action.

FAQ

What data leaves my network?

The only data that leaves your network is a BLAKE3 hash (32 bytes) submitted to Sigstore Rekor for timestamping. No PHI, no clinical content, no patient identifiers, no message payloads. The hash is a one-way cryptographic commitment that proves the chain existed at a point in time.