Building AI-native software for healthcare. Mostly databases, agents, and the unglamorous plumbing.
CoralEHR is an EHR for behavioral health clinicians, running in production. I own the AI documentation agent, the billing engine, and the compliance code that stops either of them from doing something regrettable.
BonfireDB is an open-source clinical backend where governance lives in the database instead of a middleware layer everyone forgets to call. Agents can propose writes. They cannot approve their own. It passes 133/133 HL7 SQL-on-FHIR conformance tests, which I enjoyed far more than is reasonable.
Before that I owned the classification and policy engine at ArcOne: hybrid PII detection, and a policy platform built so governance teams could stop filing tickets at engineers.
- Making agent context cited and scoped, so you can point at the exact row that justified an answer
- Whether MCP tool surfaces actually help agents. Aanish built and ran the pre-registered study that settled it: yes, by 9.5 points, on 43% less payload. I got to be one of the approvers who sign off the sealed anchors, since the protocol very sensibly does not let you approve your own
- Hybrid retrieval that fuses vector and full-text search, then refuses to hand back anything you were never allowed to see
- Row-level security that fails closed from the very first row, because the alternative is a genuinely bad week
Python TypeScript PostgreSQL AWS FHIR MCP
I ship fast, break prod occasionally, and write the test that stops it happening twice.
More at aamerjalan.me