The Indistinguishability Benchmark
In progressA multiplayer environment that measures whether AI agents can pass as human players. An eval for agent behaviour, and the same problem from the other side as bot detection.
- Evals
- Agents
- Multiplayer
Applied AI · Platform engineering · Dallas, TX
I build AI systems that hold up in production.
Principal engineer and engineering leader with 25+ years shipping high-scale software, from healthcare platforms to games with tens of millions of players. Most recently at The Wildcard Alliance I built the studio's applied-AI systems and the real-time platform behind its live-service title. I work on the platform under the model: durable agent runtimes, evaluation people can trust, and the infrastructure that keeps both affordable.
Featured project
LiveWaypoint
A service for running LLM agents as durable jobs: runs stream to any client, pause for human approval, survive restarts and can be rewound. A research agent with web search, a critic and a citation verifier, measured by an eval harness with a calibrated judge.
- briefings with 80%+ of claims cited, after a citation verifier
- 48% → 95%
- briefings with 80%+ of claims cited, after a citation verifier
- judge agreement with human labels, after recalibration
- 92%
- judge agreement with human labels, after recalibration
- to re-test a prompt change over 24 runs, by replaying every unchanged model call
- $0.90
- to re-test a prompt change over 24 runs, by replaying every unchanged model call
- FastAPI
- LangGraph
- Postgres
- Redis
- React
- Evals
More work
synthetic-arena
In progressA Rust and Bevy multiplayer engine where human browser clients and AI agents are the same kind of participant, so the server cannot tell them apart. Netcode over WebTransport with lightyear.
- Rust
- Bevy
- lightyear
- WebTransport
Production AI work
ProductionThe applied-AI systems I built at The Wildcard Alliance: an LLM content pipeline on IBM Granite, a vision-model captioning service on ChromaDB, agents that playtest builds in CI, MCP servers in daily engineering work, and an adoption program that reached 80% of the team.
- IBM Granite
- LangChain
- ChromaDB
- MCP
- Sentry
Live-service games
ShippedWords With Friends through its peak growth, Creativerse through years of live operation, Wordscapes for tens of millions of monthly players, and a live-service UE5 title whose viewer platform sustained 10,000+ concurrent viewers per match.
- UE5
- C++
- C#
- Unity
- Redis
- MongoDB
- Kubernetes
Writing
- Measuring before fixing: raising citation quality with evals
A low groundedness score turned out to be mostly the judge. How calibrating the instrument first, then adding a citation verifier, took claim coverage from 48% to 95%.
Experience
- The Wildcard Alliance · Principal Engineer2022 – 2026
- PeopleFun · Principal Engineer2021 – 2022
- Playful Corp · Principal Engineer2013 – 2021
- Newtoy (acquired by Zynga) · Principal Engineer2010 – 2013