I build production AI for operations teams, lead the people who run it, and teach organizations to adopt it. At DoorDash: an AI support workspace other teams now run their own agents in, an outsourced support team working live cases in it, and 800+ people trained. Before any of this, I led a crisis-response team as an EMT.
Claude Code power userAnthropic's Claude Code team: “You're one of Claude Code's top users”
01 Lab
Demos that run in your browser, and two open tools for working with AI agents.
Flagship · Playable build
A first-person walkable starship in Three.js with zero asset files. Every panel, texture, and planet outside the windows is generated procedurally at runtime. Built with a fleet of Claude Code agents I directed, gated by an automated screenshot-verify pipeline.
Four titles where you can watch an AI think, in vanilla JS on raw canvas: one plays a maze game and shows its reasoning, one plays you at Four in a Row and shows its search, a flock of UFOs learns to fly, and a lander teaches itself to touch down on a mesa.
How it worksEvery game is a module on one small engine: a fixed 60-step loop, a fixed screen scaled to yours, and sound synthesized live. No libraries, no sound files, no build step.
The bar at the bottom of this page. A 4 MB model on your own device matches your question to one of 62 answers I approved. Nothing is generated, so it can't make things up.
How it worksA small embedding model turns your question into a vector, and the closest of 494 sample phrasings picks the answer. When nothing is close enough, it says so instead of guessing.
A Claude Code plugin for someone's first pull request. A skill walks them through branch, commit, pull request and merge in plain words, and a hook keeps the agent to the git commands a first pull request needs, so nothing it does can't be taken back.
How it worksThe hook reads every shell command before the agent runs it. Git runs only from a short list. The classic ways to lose work get a one-line reason and a safer command, and anything the hook can't read exactly is refused too.
An eval harness with its failures left in: a made-up field guide for trail volunteers, 100 questions with hand-checked labels, and five checks scored by code. The grounded baseline passes 26 of 40 questions it never saw, and the gallery shows the other 14.
How it worksThe guide, the rubric and the questions were committed before any code. Each version is tuned on one set and runs once on a set it hasn't seen, and only those scores get quoted. A set studied to find the next fix stops counting as held out, so a new one is written after each freeze.
02 Featured program
Support agents diagnose point-of-sale integration cases across half a dozen internal systems, graded against a long troubleshooting playbook. The copilot answers from that same playbook, other teams now run their own AI agents inside the workspace, and an outsourced team works live cases in it. This is the whole loop in one program: build it, prove it, roll it out, measure it.
Every case meant hopping across a stack of internal tools, and the answer to “what next?” lived in a long playbook agents are audited against. A generic chatbot can't answer from that.
Answer from the playbook, not the model. Retrieval runs on the server and returns the sections it used, so every grounded answer cites a real source and the model never gets to invent one.
A React and TypeScript workspace, a Node backend-for-frontend, and retrieval tuned on the real playbook instead of a test fixture, which exposed four defects the fixture had hidden. Model changes are proven through the company's AI gateway before they ship: a control, the new setting, and a request that should fail. Other teams' agents plug in as panels instead of shipping another tab.
The first outsourced support team couldn't get in, and every failure looked like the same generic error. I proved their access one layer at a time (identity, device trust, app permissions, then a blocked internal call), generated their training kit from one source with a verifier, and now coach their live cases against a resolution-time target.
DoorDash's numbers stay internal. The write-up covers how each result was measured.
03 Selected work
Each one started as friction I'd lived with or a team brought to me. I built the tool, wired the measurement, and stayed through adoption. They're internal, so you get schematics instead of screenshots. Open any card for the case study.
2025 · Program + business case
A POS partner sunset its API. I sized the loss, wrote the funding case, and ran the program that kept nearly all of the revenue, under budget.
2025–26 · Program + live workshops
From a leadership-summit workshop to an org-wide series where people from the business demo production work they shipped.
2025–26 · Full-stack product
Turned a multi-day spreadsheet workflow into a real-time product, and fixed the driver-level bugs that stall teams for weeks.
2025–26 · Production platform
One real-time view of queues, service levels and staffing across outsourced vendors, replacing a legacy SaaS tool.
2025–26 · Production AI agent · Tech lead
Agents said “wall of text, then abandon.” I rebuilt the answers into a format they act on, and other agents adopted the template.
2026 · Financial model + dashboard
The org's first model of what a unit of support work costs, in-house and at vendors, so automation gets judged against real money.
04 How I build
I give it what I'd give any team: a playbook, shared memory, the right tools, and a review before anything ships. Most of it is public, so you can check the work.
We're making it official: you're a Claude Code power user.
You're one of Claude Code's top users, and we wouldn't be here without you. Thank you for building with us.
Big builds run as a fleet of Claude agents in Claude Code. On Starship Explorer, an orchestrator wrote the design contracts, Sonnet and Opus agents built in parallel lanes on separate files, and nothing merged until a reviewer signed off on types, functional tests, real-GPU screenshots and a frame-time budget. The contracts and gates are in the repo.
Work I repeat becomes a skill: a written procedure with its own checks, so every run goes the same way. Mine brief a session, run a retro, red-team a plan, and publish this site's Hill Money Watch digest through a script that rejects any draft that breaks the contract.
Hooks do what shouldn't depend on anyone remembering: each session opens with the project's state already loaded, and every commit is logged for the retro.
Decisions, error patterns and fixes carry over between sessions, and a mistake that shows up three times becomes a written rule. I open-sourced the core of it as agent-os: hooks, skills, memory and a session lifecycle for Claude Code.
MCP servers give agents real tools instead of guesses: a browser to test in, current library docs, the data a task needs. When nothing fits, I build the server.
Model changes pass a test before they ship. On the copilot, a retrieval change had to rescue questions that used to come back empty without pushing a right answer down. The answer bar on this page is scored on questions written before any model saw them: right first 81% of the time, in the top three 93%, and a test fails if an edit lowers either. How it was evaluated.
05 Background
Most engineers don't have this path: ER scribe lead, then EMT and crisis-team lead, then operations, then production code. The pattern: every role ran the operations the next one automated.
06 Toolkit
TypeScript, React, Node, Go, Fastify, Snowflake, and Claude Code for building with AI.
Email is the fastest way to reach me.
Phoenix time · UTC−7 all year; Arizona skips daylight saving