Building small models that watch larger automated systems for failures. The interesting part isn’t the models — it’s how you test something whose whole job is noticing when things go wrong, when its own failure is invisible by definition.

An agent system that runs a live site end to end. Started as ‘I wonder if’ and got out of hand. The live site is the demonstration.

Multi-touch tracking and code-level instrumentation that survives contact with reality. Built for marketing teams that want to know what’s actually driving the numbers, not what the platforms claim.