I go deep before I build. When I pick up a problem I read the papers, trace the baselines, and find the gap. My PM2.5 forecasting work started with understanding why standard temporal models lose recall on extreme pollution events — the architecture came from that analysis, not the other way around. That research-first habit runs through everything I work on.

On the engineering side I close the gap between prototype and production. The SRE Incident Responder ran against real Redis and PostgreSQL with live state graders, not mocks. ARM-Gym put every generated assembly through three-stage hardware verification. I care about the distance between a working demo and a system that holds up under real conditions, and I close it deliberately.