I'm a software engineer and I like making systems faster, cheaper and more accessible. Usually that starts with profiling, figuring out where the time is actually going, and fixing that. I started out on data pipelines and backend services and have slowly moved down the stack into inference, compilers and GPUs.
Right now I'm at UCSD's Hao AI Lab working on FastVideo, an open source video generation framework, with 24 PRs merged so far. A lot of my recent work has been on the Cosmos Predict 2.5 models. I added the distilled versions, which only need 4 steps, and hooked them into DreamVerse so a scene keeps going from its last frame. I also got FP4 attention running on the DGX Spark. The kernels were only built for the RTX 5090, but they work on the Spark too, so the full 4-bit path runs there now. Before that, I made the sparse attention kernels 18-22% faster because the best setting was outside what the autotuner was searching.
Cosmos is a big part of why I got into world models. If a model can mirror how the world actually behaves, you can build some really cool stuff on top of it. That only works if these models are fast and cheap enough to run, which is the part I like working on.