NeoAavya Clinical Engine
Prepared 2026‑08‑23  ·  v0.1.0

Security & Data Handling Overview

What happens to a file from the moment it's uploaded, and what's actually been verified — written for the team evaluating this tool before a pilot.

Pre-certification stage — see “What we don't claim” below before relying on this document

Overview

NeoAavya Clinical Engine converts Excel-based clinical trial design specifications — forms, fields, code lists — into CDISC ODM v1.3.2 XML and Dataset-JSON. Every export is validated against the official CDISC schemas before it's returned. The engine works with trial structure, not trial subjects: it has no feature that accepts, expects, or is designed to process patient-level data.

Data handling

Two paths, disclosed honestly

Most vendors describe one data-handling story. This engine actually has two, and they behave differently — worth knowing which one a given integration uses.

Path A — Stateless

Conversion endpoint

A file is processed in memory and temporary storage, then discarded. Nothing is written to a database. Verified: cleanup runs on every exit path, including errors — not only the success case.

Path B — Persistent

Authenticated platform

Accounts, saved studies, and review workflows store the converted structural metadata. The original uploaded file is discarded once processing completes — it is not kept alongside the converted output.

On both paths, a structural content check runs on every submission and blocks anything shaped like a patient identifier before it is processed, stored, or sent to any AI provider.

Controls

Verified, not just documented

Every row below was exercised against the running system, not asserted from a diagram. The right-hand column is how.

ControlWhat it doesVerification
Tenant isolation
Database-level row security, not just application filteringUnprivileged role, unfiltered query → zero cross-tenant rows
Session revocation
Logout invalidates a token immediately, not at natural expiryLive test: rejected the instant after logout
Brute-force protection
Login attempts rate-limited per account and per source IPLive-tested against the running system
Audit trail
Append-only, hash-chained — tampering is cryptographically detectable. The same control FDA Part 11 and EU Annex 11 both call forChain integrity independently re-verified
Identifier redaction
Blocks identifier-shaped content before it reaches storage or any AI providerTested against representative identifier patterns
Upload limits
Size caps on every file-accepting endpointTested incl. corrupted/malformed files
Data minimization
Raw uploads auto-deleted once no longer neededFile confirmed present, then confirmed gone
Schema validation
Every export validated against the official CDISC schema before returnEnforced in code, negative-case tested

Candor

What we don't claim

Not SOC 2 or HITRUST certified

That process runs 6–12 months and five figures or more — premature at this stage. In its place: full transparency. Our threat model and requirements-to-test traceability matrix are available on request, and every control above is independently checkable in our test suite rather than asserted on trust.

Not a substitute for your own data governance

The system is built to reject patient-identifying content structurally, but your organization's upload discipline remains the primary control. This is stated explicitly in the data-handling warranty we ask pilot clients to accept.

No BAA program, by design

Because this product is built to never hold PHI, we haven't built a Business Associate Agreement program around subprocessors — signing one would pull the product into a heavier compliance category it isn't designed to need. Happy to discuss this position directly with your compliance team.

Not a formal Part 11 or Annex 11 compliance claim

The audit trail above is the substantive technical control both FDA 21 CFR Part 11 (US) and EU GMP Annex 11 (EU) call for, but neither is a certification granted by having the control — both need the surrounding validation, SOPs, and legal review this stage of the product hasn't done yet. We'll say exactly that if asked, not round it up to “compliant.”

Available on request

  • DOC Full threat model — 9 categories, residual risk stated explicitly for each
  • DOC Requirements-to-test traceability matrix
  • DOC Breach notification procedure
  • DOC Backup & restore procedure — tested, not just documented