MetawatchPHR

Refining Raw Medical Record Chaos into Personal Clinical Sovereignty

Comment

Inspiration

Being a Moonshots fan, I could not help but be inspired by its team of contributors. A data nut by nature, I was inspired to use my skills for the betterment of humankind.

What it does

We engineered MetawatchPHR — a fully automated, privacy-first data refinery for personal health records. It behaves as a "headless EHR": advanced agentic AI ingests decades of unstructured, scattered medical files and forges them into a unified, chronologically sorted, FHIR-modeled personal health record — one timeline, charts drawn against clinical reference ranges, SOAP-structured visit notes, and hospital imaging viewable in-app. Everything runs on the user's own computer: no cloud, no account, and records never leave the machine.

How we built it

The company itself is the build system: named AI seats — CEO, Architect, QA, Legal, Security, Publications — run on Anthropic's Claude under a written governance system, with Google's Antigravity (Gemini-powered) as the implementing engineer for production code and a local Gemma-class model (via LM Studio) as the product's runtime AI. Obsidian is the coordination backbone. The deployed product also makes real Gemini API calls as its cloud AI option — the user's own key, direct to Google, behind explicit consent.

Challenges we ran into

Drift. Despite detailed design specifications, as the project matured, drift developed — threatening to halt development and poison the client record. Locking supervisory design, program management and quality assurance to Claude, implementation to Gemini Antigravity, and record conversion to Gemma established a firm baseline — and every process failure was written down as a numbered governance finding with a rule to prevent its repeat: 138 findings distilled into 54 standing rules, at last count. Inter-agent coordination stayed aligned through Obsidian.

Accomplishments that we're proud of

A coordinated multi-AI company that shipped a real product while maintaining strict data sovereignty — records never leave the user's machine — with SOAP-structured clinical notes; recognized standards (FHIR, LOINC, SNOMED CT, UCUM); per-record provenance labels and cryptographic tamper-evidence; a live store proven with a real transaction; six bilingual (EN/FR) legal instruments shipping in every installer; and a Windows beta in external testers' hands.

What we learned

AI is amazing, but needs a creator and decision maker to succeed with a vision.

What's next for MetawatchPHR

With MetawatchPHR released, everyone can own their complete personal health record — and, with their separate, explicit consent, contribute de-identified phenotype data toward the ultimate goal: a global effort to decode human genome function by linking real phenotypes — lifetimes of structured health outcomes — with genetic profiles. Many companies sequence genomes and supply SNP analysis; the phenotype half is what has been missing, and MetawatchPHR already builds it. In a world where data is king, consent-clean access to structured health knowledge can deliver untold benefits to humankind.

Built With

Share this project:

Updates