I spent most of my career moving regulated data around safely, first for planes, then for medicine. Doctorate in deep learning, two startup exits, engineering and product orgs of about a hundred people.
Yosemite Crew - an open-source operating system for animal health. Scheduling, medical records, invoicing, inventory: the plumbing a clinic actually runs on. AGPLv3, with web, desktop and mobile apps built from one TypeScript monorepo.
openrunic - the same idea pointed at human health. Early, and public since the first commit.
Most of my code is in the YosemiteCrew org, not in this account's repo list. Start with the Yosemite-Crew commit history and the pull request reviews; I ship and review daily. Every pull request passes SonarCloud with 95 percent coverage on new code and zero new issues, plus security scanning and a full typed monorepo build, before it merges.
One housekeeping note: this account is younger than my career. I lost access to my older GitHub profiles, and the work that lived on them. If you come across my prior projects out in the community, let me know.
Most of the code here is written by AI agents, deliberately and at scale. My job is the part that survives contact with production: choosing what gets built, drawing the system boundaries, and building the machinery that catches what agents get wrong. The rules they work under are versioned in the repo itself: AGENTS.md is a hierarchy of operating instructions that Codex, Claude Code and friends load before touching a workspace.
An agent will happily merge anything that passes, so what passes is where I spend my judgment. All of it is public: PR governance and a promotion guard between dev and main, secret scanning, dependency review, supply-chain scans with SBOMs, pinned transitive dependencies checked against advisories in CI, container and infrastructure scans, an OSSF scorecard, attested releases, and a workflow audit so the CI cannot quietly drift either. Health data gets its own tripwire: CI fails when PHI turns up where it should not.
The judgment calls are on the record too. A provider-independent auth boundary before the SuperTokens migration (#1763). Stripe direct charges so the clinic, and never the platform, is the merchant of record (#1680). Tenant scope derived from resources instead of trusted from the caller (#1860). A pet passport whose clinical records carry attestation (#1675).
openrunic ships an agent of its own, so the governance ships as code: the agent loop enforces approval gating, tenant scoping, budget caps and audit emission, and the clinical safety package screens prescriptions for allergies and cross-sensitivities before a suggestion reaches a human. Agents type faster than I ever will. Not one has asked me whether the backup restores; there is a job for that too.
TypeScript end to end, in a pnpm and Turborepo monorepo. Next.js and React on the web, React Native on mobile, Electron on desktop. APIs in Express and Hono. Postgres through Prisma and Supabase with row-level security, Redis and BullMQ for queues, Socket.IO for realtime.
The clinical side speaks the standards nobody sees and everybody depends on: FHIR, HL7v2, C-CDA and CDS Hooks. Around them: Stripe Connect for payments, Firebase for push, AWS for storage and email, infrastructure as code with CDK, Docker images for self-hosting.
Tested with Jest, Playwright and Storybook. Gated by SonarCloud, CodeQL, dependency review, secret scanning and supply-chain checks. Boring on purpose.
