About

I'm a CTO who got a little obsessed with one problem: AI agents are prolific and occasionally brilliant, but you can't trust them on their own. They produce things that are plausibly correct in the narrow but often drift from the long-range context and goals of humans. So I started measuring exactly where they go wrong and building the pipeline that defines what hard guarantees we can make and what remains soft around the output.

I'm a practitioner, not an academic. I don't do research to prove things, I do it to make informed choices when deploying real systems. The papers exist because, after enough deployments, I wanted the decisions backed by measurement instead of vibes.

I spent my career in distributed systems, where the hard-won lesson was that you engineer reliability into the system instead of hoping the parts behave. That whole field was built by people bringing rigor to a messy practical problem so engineers could ship correct systems. I'm applying that same lens to AI agents: name the failure modes, measure them, and put gates where they belong.

The through-line across all of it is one idea, and naming it is the point. Distributed systems got "happens-before", a primitive that lets you reason about what's actually going on. Agent reliability needs its own. The one I'm proposing is the verification surface: reliability is a property of what you can verify against, not of the model you happened to pick. You can only trust what you can verify.