I'm a software engineer drawn to operational problems — things that are slow, manual, or breaking under load. Most of my work is the same shape: profile it, find the friction, ship a measured fix. Lately that's pulled me toward AI systems, where the same reliability and latency problems show up but the tooling is younger — and steadily further down the stack, from pipelines and services into inference, compilers, and GPU kernels.

The friction is usually something quiet — a deferred cost being misattributed, a default nobody benchmarked, a tool whose own numbers are wrong. At the Hao AI Lab I work on FastVideo, an open-source video-diffusion framework: I cut Cosmos 2.5 inference latency 15.6% after finding a "slow stage" that was really absorbing a half-gigabyte GPU→CPU transfer, built a model-agnostic caching module that takes 23% off Wan2.1 at matched quality, and ported the framework to NVIDIA's DGX Spark. Same instinct outside the lab — a Go log engine doing 42M rows/hour with sub-millisecond lookups, and an ML anomaly-detection pipeline that grades its own output.

Most recently I've been testing whether that caching work transfers out of video and into robotics — Factory Diffusion measures adaptive residual caching on a LeRobot diffusion policy, trading latency against closed-loop task success. What I enjoy is the impact side of it — making systems faster, cheaper, and more accessible.