Trust & Compliance
Built to clear your security review.
One URL to forward to your IT, security, MLR, or procurement team. Everything they need, architecture, subprocessors, certifications, the no-PHI posture, the data-handling answers, consolidated. Designed to clear a vendor security questionnaire in one read.
Architectural posture
No PHI ever crosses. By design.
Veeva CRM sits inside the client's HIPAA-compliant environment. What flows from Veeva to OVYN is already de-identified before crossing the boundary, minimum cohort of 25, no PHI, no PII. We see what performed, not who experienced it. That makes us a non-covered entity by design, no BAA needed because no PHI ever crosses.
Headline
We are not a covered entity. No BAA required.
DPA available · security review on request
This page answers
The 12 questions your security team will ask.
· Do you have SOC 2?
· Are you HIPAA compliant?
· Do we need a BAA with you?
· Where does our data live?
· Is our data used to train your AI?
· Who are your subprocessors?
· Are you HITRUST certified?
· Are you 21 CFR Part 11 compliant?
· What's your incident response SLA?
· Do you sign DPAs?
· Can we run a pen test against your system?
· Where does the GenAI inference happen?
Compliance posture · framework by framework
Where we stand on each certification or framework.
Not a HIPAA covered entityArchitectural · enforced by data flow
OVYN does not handle PHI. All Veeva CRM data is de-identified by the client before crossing the boundary, minimum cohort of 25, no patient-level records, 14 aggregate fields max. Therefore no BAA is required between the client and OVYN. We are happy to sign a DPA on request.
SOC 2 Type 2, infrastructureInherited from subprocessors
All upstream subprocessors (Anthropic, Vercel, Resend, Turso) hold or are pursuing SOC 2 Type 2 certification. Reports available on request from each provider's trust portal.
SOC 2 Type 2. OVYNRoadmap · 3–4 months for Type 1
OVYN as a company is not yet SOC 2 certified. We're a small studio. SOC 2 Type 1 process can be initiated within 30 days if the client requires it as a deal condition; Type 2 follows after a 6-month observation window.
HITRUST-alignedAligned · not certified
We follow HITRUST-aligned data-handling practices: minimum-necessary access, encryption everywhere, audit trails, breach notification SLA. Not certified, alignment, not assessment. Common posture for boutique vendors.
FDA / OPDP compliancePre-MLR rail · standing instructions
Generation rail enforces fair-balance proximity, comparative-claim restrictions, outcome-guarantee prohibitions, off-label scope, and ISI requirements before content reaches your MLR team. Every variant is scored against your standing rules with auto-rewrite suggestions on flag.
21 CFR Part 11Veeva is the regulated layer
Part 11 governs electronic records / signatures for FDA-regulated systems. Your Veeva instance is the regulated record. OVYN sits beside it; the moment content lands in PromoMats, your Part 11 controls take over. We don't claim Part 11 compliance ourselves because the regulated record-of-truth is Veeva.
Subprocessor list · disclosed up front
Every vendor in the data path. With certifications + scope.
Standard pharma vendor questionnaires ask for this, most agencies provide it grudgingly, only on request. We publish it before you ask. New subprocessors require 30 days' notice and your right to object.
Vendor
Role
Region
Certifications + scope
Trust
Anthropic
AI inference
US
SOC 2 Type 2
Receives generation prompts. Per Anthropic enterprise terms, customer prompts and completions are NOT used to train foundation models. 30-day retention max for abuse monitoring; can be disabled.
Vercel
Hosting + edge compute
US
SOC 2 Type 2 · ISO 27001 · HIPAA-eligible
Serves the OVYN application. No customer content stored at rest in Vercel, used as compute layer only. SSL/TLS terminated at edge.
Resend
Transactional email
US (EU sub-processing available on request)
SOC 2 Type 2 · GDPR/CCPA-compliant DPA available
Sends magic-link sign-in emails, Research-pillar weekly digests, and operator notifications. No patient data, no PHI. Used only when the client opts into transactional email.
Turso (libsql)
Operational database
US (us-east-1)
Encrypted at rest · audit reports on request
Stores OVYN platform metadata only, variant IDs, brief structures, user accounts. NEVER stores patient-level data. AES-256 at rest. Multi-region replication disabled by default. Provider's current attestation status available on request to trust@ephicacyhealth.com.
Data flow · where your data goes
What each vendor sees. And what stays in your control.
Pharma's most common objection to AI vendors is the fear that proprietary brand strategy ends up in a public model. This is the answer, per vendor in the data path. Plus the upgrade paths when a CISO requires stronger isolation.
AI inference
Receives
Brand brain JSON · prompt · retrieved corpus pieces. Per generation.
Retention
Zero-data-retention (ZDR) configured on the production API key. No prompt or completion stored past the request lifecycle. No vendor engineer has access to your content because the data is not retained.
Training
Contractually prohibited from training on customer inputs. Clause available under NDA.
Visual engine
Receives
Visual prompt · style references · brand visual rules. Per image.
Retention
Generated images stored on the vendor's CDN for delivery; we mirror them to our object storage on success and the vendor's copy expires per their retention policy (typically 30 days).
Training
Customer prompts not used for training per vendor terms.
Database (Turso)
Receives
Brand brain · corpus · generated outputs · audit log · user records. At rest.
Retention
Stored encrypted at rest (AES-256). Field-level encryption for sensitive columns is on the SECURITY_HARDENING_PLAN roadmap (Phase 2.2). Once shipped, only the inference call sees plaintext.
Training
Database vendor cannot read content (encrypted). No training, no access.
Email (Resend)
Receives
Magic-link sign-in emails · operator notifications · subscriber digests. Per send.
Retention
Vendor's standard email-log retention. Content is non-sensitive operational text only. No brand IP transits this path.
Training
N/A · email send vendor, not an AI vendor.
If your CISO requires stronger isolation
Bring your own AI key. Your Anthropic / Azure OpenAI / AWS Bedrock account handles inference. Your contract, your logs, your audit trail. The platform never sees raw model traffic.
Per-tenant database. Your brand brain + corpus + outputs in a dedicated database instance, your region (US / EU / APAC), your backup schedule, your KMS encryption keys.
Customer-cloud deployment. The full platform deployed as a private instance inside your Vercel team or AWS account. Same code, your infrastructure. White-label optional.
Air-gapped. Open-weights models (Llama, Mistral) on your GPU cluster. No internet egress from the generation path. Brand brain never leaves your building.
Each of these is a documented deployment mode, scoped per engagement. Contact for the security pack including pen test, SOC 2, DPA, and the mode-specific architecture diagram.
Architecture · the facts
Specifics for your IT review.
Region
All US-based hosting (us-east-1). No data egress outside the US by default.
Encryption · in transit
TLS 1.3 enforced on every connection. HSTS 1-year preload eligible.
Encryption · at rest
AES-256 across all subprocessors. Database, object store, backups.
Authentication
OAuth 2.0 to Veeva sandbox first, then production with read/write scope. No shared credentials. SCIM-ready for enterprise SSO during onboarding.
Audit log
Every variant generation, MLR decision, deployment, and Veeva sync is logged with timestamp + actor + content hash. Immutable, append-only. Exportable for audit.
Backups
Daily snapshots, 30-day retention. Cross-region copy disabled by default per data-residency policy.
Incident response
Documented runbook. < 4hr notification on confirmed breach. Quarterly tabletop exercises.
Penetration testing
Annual third-party pen test scheduled for Q3 2026. Internal continuous SAST/DAST in CI.
The question pharma asks every AI agency
"Is our data used to train your AI?"
- No client data is used to train any foundation model. Period.
- Anthropic enterprise terms prohibit it for our account; we have it in writing.
- Brand brain inputs (claims library, ISI, voice rules, performance data) are retrieved at generation time only, not absorbed into model weights.
- Outputs you approve and we deploy are also not used to train any model.
Documents available · on request
What we'll send your security team.
Data Processing Agreement (DPA)
Standard CCPA/GDPR template, signed
Subprocessor change notice
30-day advance notification policy
Vendor security questionnaire response
Pre-filled SIG Lite + CAIQ
Architecture diagram
Full data-flow diagram, no-PHI annotation
Insurance certificates
E&O + cyber liability + general
Sample audit log export
JSON-formatted, last 30 days
Subprocessor SOC 2 reports
Routed from each vendor's trust portal
BAA template (if needed later)
If scope expands to PHI handling
Send the request to: trust@ephicacyhealth.com — typical turnaround under 1 business day. We'll route the right document and offer a 30-min call with whoever your security lead wants to ask questions of.
Once your security team clears
Send a brief, first 12 versions in 48 hours, no commitment.
Forward thisOne click to the right page for whoever you need to loop in.
What's underneathArchitecture · advisors · published thinking.
Talk to us30 minutes with Maha + Zein. No deck, no sales motion. We answer your questions.