Now

Building light-search: a state-space solver that learns how actions transform states locally, then turns that local knowledge into topology at search time, instead of training a model to guess answers it never really knew.

Alongside it, the memory layer for an always-on assistant I'm building: an epistemic engine, a knowledge base and memory for an agentic system. Near term, I want it to compound how fast I learn and build.

By day I contract on a database query compiler: the language, its multi-target compiler, and a parser that treats parsing as navigation, not verification. Every byte-level state carries its full set of valid futures.