Now

I'm at NVIDIA, in AI for Media, working on human pose and motion models — and increasingly on agents. My LinkedIn headline says "poses/agents," which is about as accurate a summary as I can fit in nine characters.

Most of my energy goes to turning research models into things that survive production: C++/CUDA integration, TensorRT inference paths, multi-stream batching, and the deeply unglamorous work of building evaluation harnesses that tell you when a model is quietly getting worse. Evaluation is the part nobody puts on a slide and the part that decides whether a model ships.

The other half is LLMs — post-training, LoRA fine-tuning, adversarial evaluation of agents. Coming from computer vision, the thing that surprised me is how much of it is the same problem wearing different clothes: you still can't improve what you can't measure.

Training LLMs from scratch — GPT-2 scale, the whole thing, so the abstractions stop being magic. Reading a lot on world models and vision-language-action models; I think the interesting frontier is where video prediction and robot control stop being separate fields. More on the reading page.

I started a Substack. The plan is to write about what I actually learn at work — large-scale evaluation, inference optimization, the gap between a research checkpoint and a shipped product — plus the non-technical things I can't stop thinking about.

Reviewing for the HuMoGen workshop at CVPR 2026, and judging hackathons around the Bay — most recently the Open World Hackathon on physical AI. It's the fastest way I've found to see what people build when nobody tells them what to build. The full list is here.

Two obsessions I can't shake. Longevity — not the supplement-influencer version, the actual biology of healthspan, and how little of it is settled. And energy — specifically how far behind the US has fallen on buildout speed versus China, and why I think enormous amounts of capital are about to move into solar and nuclear.

I also play guitar badly and keep trying to make Linux do things it wasn't designed to do. Both are hobbies in the sense that they mostly produce error messages.