Post-training, synthetic data, and RL for agents that perceive and act. Meta Superintelligence Labs
I post-train multimodal models that act in real environments. Given a capability we want, I build the tasks, environments, rewards, and synthetic data that teach it, then train with RL. Right now I care most about RL environments and post-training for computer-use and real-time multimodal agents.
Now at 
Meta Superintelligence Labs on real-time omni models for personal agents. Previously Visual Intelligence at 
Apple and 
MIT-IBM Watson on synthetic data. Ph.D. in Computer Science at 
MIT; B.A. in Physics at 
UC Berkeley. I read, write, and hike.
An omni model that perceives and responds in real time instead of turn by turn, and can be interrupted mid-action. I lead efforts in synthetic data and RL, and work on the distillation that makes it fast and natural enough to talk to.
Built the agentic long-form video generation workflow and the DPO and offline RL system behind Vibes, Meta AI's personalized video feed, and led the human-evaluation data behind it.
Built the datasets for on-device visual question answering on iPhone with privacy-preserving VLMs, and post-trained Apple's foundation model for proactive question answering on egocentric video. Shipped in Apple Intelligence at WWDC 2025.
Efficient on-device task planning and decomposition, trained on a synthetic planning dataset with a multi-LoRA post-training pipeline.
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 Spark 1.1 released, a multimodal reasoning model for agentic tasks, debuting alongside the Meta Model API public preview. Read more
Muse Image launched, the first image generation model from Meta Superintelligence Labs — agentic generation paired with Muse Spark. Read more
The race quietly stopped being about how smart models are and became about how cheaply smart they are. That changes where all the money goes.
Machines compute. Humans commit. The difference is not intelligence — it is what you are willing to wager when the evidence runs out.
Hierarchical multi-agent systems import human coordination constraints into a substrate that doesn't have them. The evidence, the mechanism, and the alternative.
