Engineering philosophy
Systems over demos.
I do not ship agents that merely look intelligent. I ship systems that can be measured, gated, and proven.
Most AI products fail in the gap between a convincing transcript and a consequential side effect. That gap is an engineering problem. It is not a prompt problem.
Evaluation is not a report you generate after launch. It is the first infrastructure you build. If you cannot fail a pull request when an agent regresses, you do not have quality — you have anecdotes.
Open source is how I hold my own work accountable. A protocol that cannot be inspected is a story. A harness that cannot run offline is a demo. I publish the tests.
Agent safety is not a policy PDF. It is runtime discipline: who may approve, what exactly was approved, whether the world still matches, whether the effect happened once, and whether an independent observer agrees.
Production quality means fail closed, measure honestly, and refuse vanity metrics. I would rather show a real failure than a green dashboard that learned to hide.
01
Reliable AI systems
Reliability is the product. Latency, cost, and fluency are constraints. If the system cannot be trusted under retry, timeout, and partial failure, it is not ready to act.
02
Evaluation-first engineering
Golden cases, baselines, and CI gates come before the demo reel. AgentEval exists because I refuse to treat probabilistic software as untestable.
03
Open source as proof
The strongest credibility is a public test suite, a PyPI package, and a threat model written in the same voice as the code. Claims without artifacts are marketing.
04
Agent safety at the effect boundary
Permission to call a tool is not permission to change the world. Safety lives in sealed intents, TOCTOU checks, exactly-once execution, and independent verification.
05
Production quality, not portfolio polish
Coverage gates, strict types, crash recovery, and honest limitations. If it is experimental, the README says so. If it is irreversible, the protocol says so.
06
Systems over demos
A demo is a path. A system is a state machine with invariants. I build the latter: evaluation loops, trust protocols, and agent runtimes that survive contact with reality.