Find out what your first health system's security review will find, before they send the questionnaire.

Where this usually goes wrong

You have an LLM feature in pilot and a group practice, health system or payer that wants it. Somewhere between the demo and the contract, their security team sends the 80-question questionnaire and asks for a BAA, and "where does the data live and who has access" has to be answered in writing, truthfully. For most teams under five engineers the honest answer includes a logging tool, a dev tool or a model provider setting that nobody put under a BAA, and finding that out during the review costs a quarter, or the deal.

What you get

One week, read-only, in your repo, your cloud console and your vendor list. I never see a patient record. At the end, a written report that says plainly where PHI actually flows, logs, model provider and dev tools included; which hops have a BAA and which don't; which of the fifteen questions health systems now ask about AI you can truthfully answer today, and the compensating control and date that buyers accept where the answer is no; how the handoff to a human and the crisis path are built and whether they actually fire; what will silently break at the next prompt edit or model deprecation, and the small golden set that would catch it; and the fix list in order, each item marked weekend, sprint or rebuild. It's usually not the rebuild. The report is written to be handed to a buyer.

The report's sections

1. Where the data actually lives
2. Your BAA chain: who is covered, who is not
3. The fifteen questions, and what you can say today
4. What happens when the AI is unsure
5. What breaks when the model changes
6. Will you find out from a user complaint?
7. The fix list, in order
8. What to say when you don’t have it

How it works

1. An intro from someone we both know, or an email.
2. A 45-minute call with no access. You show me what you want to: an architecture sketch, a screenshare, your vendor list. You leave with your BAA-chain map.
3. A mutual NDA.
4. A named, read-only login you can revoke, scoped to the repo, infrastructure code and tracing, for the review week. Revoked on day five with a written note. A sanitized fork or staging account is fine, and the report will say where that limited the findings. I never ask you to connect your real repo to anything.
5. The written report and a 60-minute readout within a week.

Who's doing the review

What I'm not selling

I don't sell a platform, a certificate or a remediation product. I make my living on paid engineering work. Some reviews lead to that, most don't, and the report is yours either way.