I'm a research intern at
Applied Intuition,
working on
for robotics. Most of my work spans representation learning, dynamics, and general pretraining for these models.
World-action models are built on video generation backbones and jointly predict future frames and robot actions.
On the representation side, I study what the video-pretrained backbone encodes about the physical world.
On the dynamics side, I measure whether the futures it predicts are accurate enough to plan and act on.
I'm also studying CS at
UC San Diego,
where I work with the
and contribute to
.
At the
Hao AI Lab,
I study how video and world models learn causal dynamics and physical principles.
FastVideo
is an open-source framework for accelerating large-scale video generation through efficient post-training, distillation, and serving.
At
Tesla Autopilot,
I worked on 3D foundation models, GPU kernels, and model training performance for FSD and Optimus.
At
Meta,
I helped build the runtime software stack for MTIA, Meta's in-house training and inference chips, to accelerate PyTorch ops for recommendation and ads models at scale.
