Bay 0410 projects · 7 articles

Software Engineering.

The foundation under everything else.

Every AI system I ship sits in software built to last: append only records, strict typing, tests that run before a deploy, and runbooks.

Scroll to walk the bay

How I work

  • Quotes that never drift. Every saved line keeps its price, tier, label and price book version, so a later price change never reaches it.
  • Team standards, solo. Strict typing and golden tests, because when I'm the only engineer, no one else catches it.
  • Safe to run twice. Idempotency at every boundary, so a replay never duplicates anything.
  • Read only, by test. The capacity planner has 0 write methods toward Denticon, held there by a test that fails the build.
  • Golden against the old system. The planner's engine reproduces 360 of 360 day cells from the workbook it replaced.
  • One key holder. One service holds the Denticon credentials, and the apps that need it read the mirror through a registry.

What I bring

Schemas with invariants, test pyramids, infrastructure as code, and systems that survive their author.

  • Invariants in the codeCaptured prices stored with the quote, append only ledgers, revision snapshots.
  • Tests that hold the lineGolden tests against the system being replaced, and type tests that fail the build on a forbidden method.
  • Idempotency everywhereWebhooks, ledger appends, uploads, ETL runs: replay anything, duplicate nothing.
  • Built to be leftTerraform, runbooks, ADRs and admin panels, so the system doesn't need its author.
In production

Patient Dashboard

Two Cloud Run services from one image: a dashboard with no Denticon credentials and a sync that mirrors Denticon into Firestore.

  • Only query parameters proven to change totalCount are allowed; the query builder throws on anything else
  • The freshness report starts from the cursors it expects, so a missing one shows as the worst status
  • Link writes are transactions that decide again on fresh reads, and every link is versioned, audited and undoable
In production

Capacity Planner

A pnpm monorepo: pure TypeScript engine, Postgres on Cloud Run, read only source access enforced by a type test.

  • 360 of 360 day cells from the old workbook reproduced on every test run
  • A mass cancel guard: a sync that would cancel an office's bookings in bulk changes nothing
  • The logger records only the Postgres error code, constraint and table, so no SQL parameter reaches a log
In production

Treatment Estimator

The chairside pricing tool: v1.0.0 in 14.5 hours, and a saved price never moves.

  • Frozen prices: every saved line keeps its captured price, tier, label and price book version
  • Per office price books that only an admin can publish, so pricing changes need no developer
  • Presented versions frozen as snapshots; reopening moves rows into the next revision
In production

Treatment Sequence

Turns an accepted estimate into a visit plan from one of eleven practice masters, exported to Word and PDF.

  • 0 errors in 271 generation attempts across every live production estimate, audit of Aug 27
  • A living plan plus immutable version snapshots, with a unique index as the concurrency guard
  • Shipped as a committed build snapshot: no registry, no token to expire, and a check that core and ui share one commit
In production

Hybridge Forms

Our own lead forms for the practice websites, built so no submission is ever lost.

  • Firestore first, a private Cloud Storage bucket as the fallback, and the browser retries with the same id
  • The logger drops patient fields by name and by shape, so no answer reaches a log line
  • One Hybridge look, delivery to the Monday board and email copies; live since October 6
Offline

Accounting automation platform

FastAPI behind Cloud IAP: upload a file, answer one form, run the script, download the output.

  • One shared contract for every script, with the first one tested end to end from upload to download
  • In browser script editing that commits to GitHub, so fixes ship without a terminal
  • Keyless CI/CD through Workload Identity Federation
In production

NPC pipeline engineering

Ports and adapters in Python: the model returns integers and gate answers, and plain code does the math.

  • Each stage behind a port and written test first
  • A run lease so overlapping runs never grade or email a call twice
Prototype

Doctor Portal

The lab's own doctor portal, where the Rx joins the case the scanner made, so no doctor sees a duplicate. Live on sample data.

  • Top priority to a working live link in 3.7 hours
  • Every release note lists what is still pretend; real data waits on a vendor license
Open source

JobPilot

My open source job hunt autopilot that watched 163 company boards, and the one project that spans all four disciplines.

  • Open sourced under MIT with a fork guide written to hand to an AI coding agent
  • Tailors a resume and cover letter per match, checked by a calibrated judge
  • Ran 1,008 times from June 10 to October 8; a public demo runs on fictional data
Live

This website

Astro and Tailwind, a hand built design system, React islands and a free tier chatbot, deployed by GitHub Actions.