Security & privacy
Designed for privacy-sensitive biomedical workflows
Privacy controls are embedded throughout the data lifecycle — from ingestion through AI dispatch — because in biomedical settings, a leak after analysis is still a leak.
The privacy pipeline
- 1
Detection
PII/PHI found via regex, dictionaries, context rules, and NER across names, IDs, dates, locations, providers, organizations.
- 2
Consistent pseudonymization
Secret-keyed HMAC maps every surface form of a patient (“John Smith”, “J. Smith”) to one stable synthetic identity; different patients never collide; tenants never share identities.
- 3
Patient-consistent date shifting
One cryptographically random offset per patient stored only in an encrypted vault — ordering and intervals preserved, partial dates keep granularity, relative expressions rewritten against the shifted reference.
- 4
Redaction & generalization
Identifiers redacted, locations generalized, document metadata stripped.
- 5
Independent residual scan
A second detector instance re-scans output. High-confidence PHI residue BLOCKS the document: empty output, never a warning. Reviewers resolve blocks through a human workbench whose corrections become tenant rules.
Documents are untrusted text: prompt-injection content hidden inside clinical documents is surfaced transparently and never forwarded to models as instructions.
Defense-in-depth
Layer 1 · Authentication
Bearer JWT via your identity provider (OIDC/SAML with MFA in production) or hashed API keys for service principals. Secrets displayed once; revocation instant.
Layer 2 · Authorization
Role-based access (admin, clinician, nurse, researcher, service, finance roles) with wildcard permission matching, plus attribute-based policy rules per tenant. Every denial is audited.
Layer 3 · Tenant isolation
Hard boundary — tenancy derived server-side from credentials on every request; patient identifiers namespaced per tenant; cross-tenant access denied even for platform operators.
Layer 4 · Purpose limitation
Data use gated by declared purpose: treatment by default, research requires consent flags, marketing is denied outright.
Layer 5 · Classification ladder
Public / internal / confidential / restricted clearance levels enforced on sensitive reads.
Layer 6 · Audit trail
Append-only, SHA-256 hash-chained records — who, what, when, outcome, request ID — with verification that reports the exact index of any tampered entry.
Layer 7 · Privacy boundary
The de-identification pipeline itself: pseudonymization, date shifting, residual scanning that fails closed.
Governed AI routing
Model selection follows a safety ladder: deterministic NLP first, locally deployed open models by default, external premium models only where entitled, opted-in, and operating on minimized PHI. In BYOC and on-premises deployments the external tier is disabled entirely — the platform enforces this server-side, not as a configuration suggestion.
Deployment modes
| Mode | Description | Notes |
|---|---|---|
| Single-tenant cloud | Dedicated environment per customer | Default |
| Customer cloud (BYOC) | Runs inside your AWS account under your governance | Gateway honors trust boundaries |
| On-premises / private AI | Institution-hosted inference with no external egress | External LLM routing disabled at the platform level |
We describe controls, not certifications: BioMedora makes no regulatory-compliance claims on this page. Architecture documentation, audit samples, and control mappings are shared during enterprise evaluation under NDA.