Training real-time multimodal agents. Research Scientist, Meta Superintelligence Labs · 2025–
I train real-time multimodal agents that talk, listen, and act while the interaction is still going. I build the tasks, synthetic data, and rewards, then train with RL.
At 
Meta Superintelligence Labs I lead synthetic data for Muse Realtime Avatar, and developed offline RL for Vibes’ personalized feed. Before that, I worked on Visual Intelligence at 
Apple and foundation models at 
MIT-IBM Watson. I got my Ph.D. in Computer Science at 
MIT, and B.A. in Physics at 
UC Berkeley. Outside work I read, write, bike, and hike.
A live, expressive avatar that talks with you in real time. I lead its synthetic data, and applied self-forcing and DMD distillation to make it fast enough to run live.
Built the launch video-generation workflow, then developed preference learning and offline RL for Vibes, Meta AI's personalized video feed.
Built the datasets for on-device visual question answering on iPhone with privacy-preserving VLMs, and post-trained Apple's foundation model for Visual Intelligence. Shipped in Apple Intelligence at WWDC 2025.
Given a target capability, synthesize the tasks, the environment, and the reward, then train with RL. No labeled dataset required; the task definition is the input.
JetMoE-8B, an open-source mixture-of-experts model trained for a fraction of the usual cost, with a synthetic corpus for pre-training and post-training. Its chat model beats Llama2-13B-Chat on MT-Bench.
Efficient on-device task planning and decomposition, trained on a synthetic planning dataset with a multi-LoRA post-training pipeline.
All publications → · Google Scholar →
At COLM 2026 in San Francisco, Oct 6–9. Email me to meet at the poster sessions or the Lifelong Agents workshop.
Read more →Muse Realtime Voice and Avatar announced at the Meta Connect 2026 keynote: one streaming system for conversation, speech, and a live, expressive avatar. I lead synthetic data for Muse Realtime Avatar.
Read more →Muse Spark 1.1 released, a multimodal reasoning model for agentic tasks, alongside the Meta Model API public preview.
Read more →A clip model never faces consequences. A real-time model does: any frame can change what the user does next. That makes it a policy, not a generator.
When the headline metric becomes output per joule, a technology has become a utility. Intelligence just did, and the money will go where it always goes: upstairs.
Machines calculate. Humans commit. The gap is not intelligence — it is what you are willing to wager once the evidence runs out.
