I am currently a software engineer at Google working on AI Infra. Besides, I did research in AI and robotics. I love taking ideas the whole way from research to a real system in the world, solving complex problems efficiently, and finally turns into a product.
Generated video and audio jointly enable force-aware robot manipulation, using contact sound as a force signal that video alone cannot capture.
under reviewUsing a digital-twin representation between surgical video and the network makes phase recognition far more robust to out-of-distribution and corrupted inputs.
International Workshop on Digital Twin for HealthcareA paired clean/corrupted benchmark built by robotic replay that measures how surgical tool segmentation degrades under bleeding, smoke, and low light.
MICCAI EndoVis'24 ChallengeDeveloped a RL policy for high-speed autonomous drone racing and overcome the sim2real bottleneck. It took 1st place in the sim-to-real track and placed top 3 in both simulation and sim-to-real among 30 teams.
Sim-to-real drone racing run · ~19s
Built end-to-end as the engineer and product owner: an AI-native marketplace that helps students monetize their skills and benefit one another. I took it from a single idea to a live product — designing AI agents, AI search & recommendation, real-time chat, and the full cloud infrastructure on Azure (database, DevOps).
An extension of edit-friendly DDPM inversion that brings ControlNet guidance into the inversion loop, giving text-driven image edits tighter structural control while preserving the original content.
Reinforcement & imitation learning, policy learning, and sim-to-real — bridging perception, control, and embodiment so models act reliably in the real world.
LLMs, MLLMs, and foundation models across text, vision, and action — reasoning, tool use, and retrieval grounded in what models see and do.
Designing autonomous, tool-using agents — planning, memory, and orchestration that turn models into systems that get real work done.
Image & video generation and diffusion models — synthesis and creative pipelines.
Building products end to end with an emphasis on high availability, scalability, and sound system design — from architecture to a shipped product.