Security & data handling.

Last updated: 24 April 2026

We're an early-stage firm. We don't have SOC 2 yet. Here's exactly what we do — and don't do — with your data, written plainly so you can decide if it's enough for your security review.

1. What we touch

• Provider invoices & usage exports — OpenAI, Anthropic, AWS Bedrock, Google Vertex AI, Azure OpenAI billing CSVs/JSON.
• Gateway logs — LiteLLM, Helicone, Langfuse, OpenRouter, or your own proxy. Aggregated metadata only (model, token counts, latency, request volume).
• Read-only API keys with billing-export scope, never write keys, never production model keys.

2. What we don't touch

• Prompt content, model output, and end-user PII are not stored in our systems by default.
• If a specific optimization (e.g. semantic cache tuning) needs prompt samples, we ask explicit per-feature approval and scope sampling tightly. You can revoke any time.
• We never use your data to train any model. We never sell or share with third parties.

3. How we handle credentials

• Credentials live in a password manager with hardware-key-protected access. They are never committed to git, never pasted into Slack/email, never embedded in reports.
• You can rotate or revoke at any time without notice. We log every credential use to your engagement audit trail.
• Engagement ends → credentials are revoked by you and purged from our manager within 7 days.

4. Where data lives

• Encrypted databases (AES-256 at rest, TLS 1.2+ in transit). Region selectable EU or US at engagement kickoff.
• Backups are encrypted and retained 30 days.
• Access is limited to operators assigned to your engagement; access is logged.

5. Sub-processors

We disclose every sub-processor in the DPA. Current list (subject to change with notice):

6. Compliance

• GDPR / UK GDPR / CCPA — DPA available on request. EU/UK transfers covered by SCCs.
• HIPAA — not currently set up. Don't send us PHI.
• SOC 2 Type II — on roadmap. We can share our controls memo and pen-test summary under NDA in the meantime.

7. Reporting a vulnerability

Email [email protected]. We acknowledge within 2 business days. See also /.well-known/security.txt.

8. Incident response

Security incidents involving customer data are disclosed to affected customers within 72 hours of confirmation. Post-mortems are shared within 14 days.