Open to remote roles · US

AI engineering · Agents · Evals · Developer tooling

I build AI systems that show their work.

Agents, evals and developer tooling, built by someone who spent four years making sure macOS, Windows, iOS and Android releases were right before they shipped. Everything below comes with receipts.

01 · Selected work

Release gates, evals and agents, each with public evidence.

OrderProof Android release gate

Checkout showed one confirmation. The ledger recorded two payments. The gate catches what the UI never shows.

UI confirmation
1 ✓
ledger charges
2 ✗
verdict
FAIL · DUPLICATE_ORDER

The faulty build fails on 2 emulator configurations, and the fixed build passes all 8 scenarios.

  • Kotlin
  • Android
  • Emulator CI

Citebound Grounded RAG

An RFC assistant that checks every quote against pinned RFC text and abstains when the source doesn't support an answer.

quotes verbatim
203 / 206
unanswerable → abstained
20 / 20

Both measured in CI. The limits are published.

  • Go
  • RAG
  • Evals

triagewright Public soon

A support agent that looks things up, cites its sources, and asks a human before it moves money. Evals grade the tool-call log, not what the agent claims it did.

tests
854 passing
eval pass rate
70.4% → target 90%
refunds without a human
0
  • Python
  • MCP
  • RAG
  • OpenTelemetry

factorysim Rust/Wasm simulator

Wire spec writers, coding agents, reviewers and CI together on a canvas, then see where the queue piles up. Deterministic, with no LLMs.

median lead time · before
12.1 h
+ one reviewer
4.7 h (−61%)
  • Rust
  • Wasm
  • TypeScript
The Cyclorama emulator running a demo cartridge next to its debugger

Cyclorama SDK Range project

A 16-bit console from 1987 that never existed, with a real toolchain: emulator, CYASM assembler, SCRIM compiler and tutorials. The machine is defined by a conformance suite of 15 ROMs.

  • TypeScript
  • Emulation
  • Conformance testing

02 · Case study

One task, start to finish, from the Factory's own records.

  1. BriefBooking flow + mugshot. The worker gets its own worktree.claude-opus · t31
  2. Check failed1 of 26 tests red. The real output goes back.EXIT 1 · 57.5 s
  3. Check passedThe exit code decides, not an agent.EXIT 0 · 58.8 s
  4. Review: changesA GPT reviewer flags mirrored JPEGs.PhotoLoader.kt:28
  5. Fix, with proofAll 8 EXIF orientations, plus a test that fails on the old code.test red → green
  6. MergedRebased, re-checked, fast-forwarded.3de05fb · r2 approve

Read the full task →

Where this came from: four years of signing off releases taught me that a green CI run and a working feature are not the same thing. Lookout and the Factory are my own projects, built from scratch on my own time around that lesson: agents do the legwork, code enforces the rules, a person decides. Watch Lookout catch a bug CI missed →

03 · How I work

Rules I enforce in code, not in prompts.

04 · Playground

Things I built for fun.