Nartis
An AI-native platform for HEDIS and CMS Star Ratings: clinical data into a governed warehouse, measure logic that is transparent and auditable, results that quality leaders can act on. Designed and prototyped.
Problem
Healthcare quality programs compute hundreds of measures across millions of patient records and report them to several regulators, in different formats, on different schedules. Most of the tooling predates modern data warehousing and treats each reporting cycle as a one-off extract.
The result is a domain that still runs on spreadsheets, where a change to a measure takes a quarter to flow through scoring, reporting and audit, and where nobody can show exactly how a number was produced.
Approach
- 01
Treat quality measurement as a continuous pipeline, not a reporting event: clinical data lands in a governed warehouse once, and measures run against it on a schedule with full lineage.
- 02
Make the measure logic readable. Every HEDIS and Stars measure is expressed in a form a quality analyst can inspect and an auditor can trace back to the source rows.
- 03
Make AI-assisted measure development native rather than bolted on, so a proposed change to a measure can be drafted, tested against historical data and promoted through scoring, reporting and audit in one motion.
- 04
Build the prototype end to end, with a whitepaper and an explainer, so the concept could be tested with people who run quality programs before any production investment.
Architecture
Clinical sources feed a governed warehouse; the measure engine runs transparent HEDIS and Stars logic with lineage; AI-assisted development sits beside the engine, not inside the numbers; reporting and the audit trail come out of the same run.
How it's built, at nartis.health/build
Built with
- Governed data warehouse
- HEDIS and CMS Star Ratings measure engine
- AI-assisted measure development
- CMS and NCQA reporting
- Lineage and audit trail
- Next.js prototype
Outcome
A working prototype, a whitepaper and an explainer at nartis.health. The platform is designed and prototyped; it is not in production.
The architecture and build notes are published at nartis.health/build.
Decisions
A warehouse with lineage over per-cycle extracts
because the audit question is always 'show me how you got this number', and extracts cannot answer it.
AI beside the measure engine over AI inside the scoring
because a regulator has to be able to read the logic; a model proposing changes is fine, a model producing scores is not.
Carplow
Watches Facebook Marketplace for each saved search, grades every listing with Claude against market value, and emails the grades you pick. Live for a small group of friends: Next.js on Vercel, AWS CDK back end.