The Lab
Four things run in public: a trusted framework for putting AI agents on health data, an agent built on top of it, an open corpus of the quality measures the industry keeps rewriting from scratch, and an audit of the federal provider directory.
All of it is open source and usable without talking to anyone. Consulting buys the time to build these; it is not what they are for. Everything here has a live URL, and if a project gets archived it comes off the page.
Projects
A trusted framework for putting AI agents on health data. Redaction, audit and human sign-off are enforced server-side, so an agent cannot read PHI unredacted or write to a chart without an approval step.
What you get
A policy broker you can run in front of your own FHIR server, rather than a set of rules you ask a model to follow.
A personal health agent you can stand up in under a minute, integrated and guardrailed by HealthClaw. Every read redacted, every access audited, every action approved by the person whose data it is.
What you get
Somewhere to watch the guardrails work end to end before you commit to building on them.
An open corpus of quality measures, and a shared record of what implementers have learned about each one. CQL and SQL next to the measure instead of buried in a specification PDF.
What you get
52 CMS eCQM packages for the 2026 reporting year with provenance verified in CI, an oq CLI that validates and packs them, and three conformance levels so a shared measure is distinguishable from a verified one. Open licences only, checked against an SPDX allowlist; value sets are referenced rather than embedded so the licensing stays machine-enforceable.
An audit of how accurate the federal provider directory actually is, scored across every state and specialty on six dimensions: field completeness, cross-source agreement, currency, endpoint reachability, federal integrity and specialty validity. Scoring is mapped to the REAL Health Providers Act.
What you get
Per-provider verification against primary sources — click through to check an NPI against the NPPES registry, the OIG exclusion list and SAM.gov, with no login in the way.
The aggregate visualisation currently runs on a deterministic synthetic seed while the measurement pipeline is wired up, and the site says so on the page rather than letting anyone cite a number that has not been measured yet.
More
Earlier experiments have been archived rather than left running as broken links. The full history, including what did not work, is on GitHub.