Q BRIDGEQ BRIDGE

DeviceBridge · Security & Compliance

Security & compliance, built for regulated submissions.

DeviceBridge files real FDA 510(k) submissions to CDRH. Every artifact a specialist agent drafts, every signature a named regulatory lead applies, and every regulatory citation resolves into one tamper-evident substrate — built so an FDA inspector or a vendor-risk reviewer could verify it remotely. This page states what is shipped and what is scoped, without overclaiming.

100%
actions captured in the hash-chained audit
9
signed gates · WebAuthn 2FA on every signature
US
data residency · US AWS regions
0
customer data used to train models
Written for the questions a CISO, a vendor-risk reviewer, and a 21 CFR Part 11 owner ask
21 CFR Part 11WebAuthn / FIDO2RLS multi-tenancySOC 2-alignedISO 27001 underwayHIPAA-ready

21 CFR Part 11 posture

The audit trail is the product's spine.

Part 11 §11.10(e) requires secure, computer-generated, time-stamped audit trails. DeviceBridge implements that as a hash-chained ledger with immutability enforced at the database tier and a standalone verifier — so the integrity claim is mechanical, not a matter of trust.

Tamper-evident, not just append-only

A per-org SHA-256 hash chain anchored to genesis.

Every action — login, gate signature, agent run, document mutation, payment reconciliation — appends to a hash-chained ledger. Each event embeds the SHA-256 of the prior event, so the chain is its own integrity proof. Append-only logging tells you a row was added; a hash chain tells you whether any earlier row was altered after the fact.

Immutability at the database tier

UPDATE / DELETE / TRUNCATE are blocked on the audit table.

The audit ledger is protected by database triggers that reject UPDATE, DELETE, and TRUNCATE outright — not by application convention that a stray migration could bypass. In the AWS healthcare-grade target, S3 Object Lock in Compliance mode extends the same guarantee to exported evidence for long-horizon regulated retention.

Standalone verifier

It reports exactly which link broke.

A verifier walks the chain on demand and re-computes every hash. If a single byte of a single event was changed, it names the broken link rather than returning a vague pass/fail. This is the artifact an inspector spot-checks: the chain can be re-verified independently of the running application.

E-signature manifests

Every signature is bound to the artifact hash it approved.

Gate sign-offs produce a signed manifest recording signer, role, UTC timestamp, intent string ("approved" / "reviewed" / "released"), and the SHA-256 of the exact artifact reviewed — plus the corresponding audit event. A signature cannot be silently re-pointed at a different document version; the hash binding makes substitution detectable.

Identity & signatures

WebAuthn 2FA on every gate signature.

A 510(k) program advances only through nine human-in-the-loop Gates. None of them can be signed with a password alone.

Two distinct factors · Part 11 §11.200(a)

Something you know, plus a phishing-resistant key.

Login is bearer-token with a mandatory FIDO2/WebAuthn second factor once enrolled. Every Gate signature additionally requires the e-signature PIN and a fresh WebAuthn assertion bound to the artifact hash. The assertion is replay-resistant: a captured signature cannot be reused against a different artifact or a later session.

  • Roaming security keys and platform biometrics (Touch ID, Windows Hello) both supported
  • Per-tenant SSO via SAML / OIDC (Okta, Azure AD, Google Workspace)
  • Wet-ink scan fallback writes the same manifest schema — never a side door around the record
  • Signature intent is explicit: approved / reviewed / released, captured per gate

Nine gates · named approvers

No agent submits to FDA without nine signatures.

Pathway and CDRH-vs-CBER routing, predicates, the substantial-equivalence narrative, the test plan, the section bundle, eSTAR assembly, user-fee reconciliation, and the final release to the CDRH Portal are each a named human's signed decision. Critical gates require multiple signers.

The same signature substrate powers INDBridge, the sister IND platform — so a sponsor running a combination product or companion diagnostic gets one consistent signature and audit model across drug and device.

See the nine Gates in detail →

Multi-tenant isolation

One platform, hard tenant boundaries.

Isolation is enforced at three tiers — database, application, and storage — so no single mistake collapses the boundary between sponsors.

Database tier

Row-Level Security on every tenant table.

Tenancy is enforced by Postgres Row-Level Security keyed on org_id, evaluated by the database on every read and write. A query that omits the tenant predicate returns nothing rather than another tenant's rows. Service-role credentials never leave the backend.

Application tier

Defense in depth, not a single gate.

The API resolves and pins the tenant from the authenticated session and scopes every query independently of the database policy. RLS and the application layer must agree; neither is the sole control. A bug in one is caught by the other.

Storage tier

Per-tenant object scoping for attachments.

eSTAR attachments, signed manifests, and assembled PDFs are namespaced per tenant. Every attachment is SHA-256 hashed at rest so its integrity can be re-verified, and the hash is what signatures and audit events reference.

The LLM boundary

PHI and PII are scrubbed before they reach the model.

The hardest question a CISO asks an AI-first vendor is what crosses the model boundary. DeviceBridge treats that boundary as a control point, not an afterthought.

Minimize at the edge

Scrubbing at the prompt boundary.

Patient identifiers and personal data are scrubbed at the LLM boundary — agents reason over de-identified, minimized inputs. A 510(k) is primarily device, predicate, and test data; where clinical content appears, it is reduced to what the section actually requires before any prompt is constructed.

  • Personal / patient identifiers scoped to the records that require them
  • Retention boundaries and deletion paths defined per data class
  • Data Subject Access Requests are first-class operations, not a manual export
  • No customer data is used to train models

Provenance over the boundary

Every model call is logged and reconstructable.

Each agent run records its tool-use trace and token telemetry. An inspector can reconstruct exactly what each agent reasoned over and which grounded sources it cited. The source-of-truth doctrine means agents refuse to emit a regulatory claim they cannot ground in a snapshotted FDA source — there is no path for an ungrounded K-number or predicate to enter a submission.

Read about AI provenance →

Data residency & hosting

US AWS regions, healthcare-grade by Day 30.

Production today runs on SOC 2 Type II sub-processors. The AWS healthcare-grade migration is a committed Day-30 implementation milestone, stated plainly rather than implied as already complete.

Current production

Vercel + Fly.io + Supabase.

Web edge on Vercel, FastAPI and the agent fleet on Fly.io (region IAD), and Postgres on Supabase Pro with Row-Level Security on every tenant table. All three are SOC 2 Type II sub-processors, documented per region.

AWS healthcare-grade · Day 30 target

US-region Aurora, KMS, Object Lock.

  • Aurora PostgreSQL Multi-AZ — us-east-1 primary, us-east-2 backup, us-west-2 DR
  • ECS Fargate compute; S3 with KMS customer-managed keys
  • S3 Object Lock (Compliance mode) for long-horizon regulated retention
  • AWS Shield Advanced; AWS Audit Manager for continuous evidence
  • AES-256 at rest, TLS 1.3 in transit

Compliance program · honest status

Aligned today. Certified on a published timeline.

Regulated buyers reject overclaiming instantly. So every certification claim here carries its real status: what is in place, and what is scoped for when.

SOC 2

SOC 2-aligned controls. Audit in progress.

The control environment is built to the SOC 2 Trust Services Criteria today. The Type II audit window is in progress (2026). We say "SOC 2-aligned," not "SOC 2 certified," until the report is issued.

ISO 27001

An ISMS program underway. Not yet certified.

An ISO 27001 information-security management program is underway, with Stage 1 scoped for 2026. We do not claim certification before the certificate exists.

HIPAA

HIPAA-ready architecture. BAA at Day 30.

The architecture is HIPAA-ready — encryption, access control, audit, and minimization are in place. The AWS Business Associate Agreement executes at Day 30 of implementation, when the healthcare-grade environment stands up. It is a milestone, not a Day-0 claim.

Penetration testing posture

Internal suite now. External assessment commissioned.

We run an internal penetration suite and a STRIDE threat-model walk across the authentication, signature, multi-tenant, and submission-RPA surfaces. An independent external assessment is commissioned for Q3 2026. We will share scope and a summary of findings under NDA with serious evaluators.

  • Internal penetration suite across auth, signature, and tenancy surfaces
  • STRIDE threat-model walk on the submission and RPA paths
  • Selector-resilient CDRH Portal RPA with a daily drift watchdog
  • External assessment commissioned · Q3 2026
  • Dependency vulnerability scanning in CI; signed commits; required review on main

Honest disclosure

Q BRIDGE AI publishes what is shipped and what is scoped. SOC 2 Type II and ISO 27001 audits are in progress (2026). An external penetration assessment is commissioned. The first production FDA submission lands with our first customer; the platform is validated end-to-end against a mock-FDA fixture and a development CDRH simulator today. Our full known-issues register is available on request.

Security review

Bring your hardest vendor-risk questions.

We will walk a CISO, a QA lead, and a vendor-risk reviewer through the audit chain, the signature model, and the data boundaries on a live workspace — and share the known-issues register under NDA.