Security & Compliance
Built for the regulatory bar healthcare AI actually has to clear.
High-risk mappings cannot auto-deploy regardless of AI confidence — a deliberate ceiling on AI authority, enforced by a risk classifier and a named human approver recorded in the audit trail.
Compliance posture
HIPAA
Designed to execute under a BAA. Every PHI touchpoint logged with actor, resource, and payload hash. Minimum-necessary enforcement via channel-level scoping.
SOC 2 Type II
Annual certification target. Compliance documentation is generated directly from the audit stream, so attestation is continuous rather than point-in-time.
GDPR
Data-subject access and erasure implemented at the channel-data-store level. PII/PHI tagged at ingestion so erasure requests run without full-table scans.
EU AI Act
Architecture designed around "high-risk" AI system obligations — documented model provenance, auditable deterministic orchestration, human-in-the-loop gates, full incident logging.
State regulations
Data-classification and audit trail satisfy stricter state regimes (CCPA/CMIA, NY SHIELD) without per-tenant customization.
ISO 27001 / HITRUST CSF
On roadmap, alongside the on-premise HITRUST CSF deployment track.
Security architecture
Identity & access
- Enterprise SSO via SAML 2.0, OIDC, LDAP/AD through Keycloak
- JWT-based authorization at every API gateway and service boundary
- Five built-in roles plus custom role builder; channel-level access scoping
- MFA required for privileged roles
Data protection
- TLS 1.2+ on every network hop, including mutual TLS service-to-service
- AES-256 encryption at rest across PostgreSQL, Redis, Kafka, S3
- PHI field-level tagging with redaction in logs and AI-agent prompts
- Key management via cloud KMS or HashiCorp Vault on-premise
LLM-specific privacy
- Zero-retention provider contracts for external LLMs
- Prompt-level PHI scrubbing before any external LLM call
- Self-hosted LLM option (Llama, Mistral) for perimeter-locked customers
- No training on customer PHI — fine-tuning is tenant-scoped and opt-in
Audit & operations
- Append-only, cryptographically chained audit trail
- SIEM integration (Splunk, Sentinel, Elastic) via standardized event streams
- Automated alerting on privilege escalation and bulk PHI access
- Signed, scanned container images with SLSA-level build attestations
Three deliberate ethical commitments
Human-in-the-loop by default
Clinical-risk mappings cannot auto-deploy regardless of AI confidence.
Confidence transparency, not marketing
Calibrated, not optimistic, confidence scores — we reject UX that encourages ignoring low-confidence warnings.
Built to be replaceable
Channels are portable because their contract is an HL7 standard, not a proprietary model — it's in our interest to keep earning the seat.