I help teams trust AI agents with real work.
I build the tools and guardrails that let teams hand real work to AI agents, and I teach people to use them well. Before that, I spent ten years building the platforms engineering teams run on.
My focus is runtime alignment: keeping AI agents doing what you actually intended while they work, within limits they can't cross and with work you can check.
Give agents access to your tools, integrations and data through connections they can't misuse. They only get the access a task needs, and a person approves the decisions that matter.
I've spent years working with sensitive student, health and financial data.
Repeatable tests of how agents and models actually perform, with cost tracked per model. Spend is broken down by model, role and phase, so you can see what each workflow costs.
Agents that check work before anyone relies on it, whether that's a contract, a research memo, a report or code. Findings are checked against the source before they reach a person, follow-ups only look at what changed, and the people using it shape how it works. The AI code reviewer I built covered every pull request at about a fifth of the estimated cost of the managed tool it replaced.
I teach teams, engineers and non-engineers alike, to work well with AI, set practical best practices, and keep the tools simple enough that they stick.
Build, test and deployment pipelines and cloud platforms that are quick, cheap to run and easy for your team to own. My self-hosted CI runners at Khan Academy cut compute costs in half.
Essays on runtime alignment: how to keep agents doing what you intended, finishing the job and staying trustworthy.
Human attention is the scarcest resource in AI-assisted work, and AI output can spend it or save it.
Keeping agents doing what you intended while they run, and the series on how to build for it. Start here.
When AI reviewers agree, that tells you their reasoning hangs together. It doesn't tell you the work is right.
Agents get useful when they reach your real systems. Design those connections so a confused or manipulated agent can't do damage through them.
I'm open to full-time and part-time roles and to consulting, especially with mission-driven teams. I'm based in South Lake Tahoe and work remotely. Tell me which workflow you want to improve, what tools it touches and where it gets stuck.