About

Most AI systems trust the model to behave. I build the ones that don't have to — where a deterministic layer, not a prompt, decides what the model is allowed to do, and a failing eval blocks the deploy before anyone notices.

Right now, that means being the sole AI engineer behind Savi, a regulated UK immigration eligibility engine at HealthCrew.ai, and shaping the AI layer of Forge, an agentic IT service desk at Kroneus Zero Trust Security. Before that, an MSc dissertation on graph-constrained retrieval cut hallucination rate 94.5% against a naive baseline — same instinct, harder problem: knowing what a model is allowed to see, not just what it's allowed to say.