Tech Founder & AI Researcher
I am the co-founder and CTO of OdyssLife, where I build always-on, multimodal agent systems that connect AI with human life. Our work is centered on turning real-world sensory input into a persistent health interface, with long-horizon understanding across vision, audio, motion, and behavioral context.
We are about to launch Odyss N1 in September 2026, an AI-powered wearable necklace designed for continuous lifestyle and dietary understanding in the physical world. It brings together sensing, on-device and cloud intelligence, and real-world behavioral modeling into a single AI-native health interface.
Before OdyssLife, I worked at Pony.ai, where I was in charge of the prediction module in the North American autonomous driving stack and was a foundational contributor to end-to-end behavioral prediction models and world models for simulation and data augmentation. That work shaped how I think about real-world AI systems: robust perception, long-horizon reasoning, and closed feedback loops between model outputs and physical environments.
Earlier, I worked on perception and computer vision problems at Waymo and Tencent Youtu. I received an M.S. in Computer Vision in 2020 from Carnegie Mellon University and a dual-major B.S. in Computer Science and Applied Mathematics in 2019 from HKUST.
My current work focuses on multimodal sensing and agent system design for AI hardware. My research interests lie in long video understanding, agentic RL, and RL training infrastructure.
We are looking for self-driven, high-caliber people who care deeply about building AI-native systems across these areas. If you think you could be a strong fit, feel free to reach out directly.
EMNLP 2025
Introduces a convolution-basis view of attention computation and uses FFT-based structure to reduce attention complexity from quadratic toward near-linear regimes for long-context transformers.
ICCV Workshops 2025
Curates a higher-quality subset of large-scale 3D data with quality annotations, enabling more reliable and efficient training for generative 3D and simulation workflows.
AAAI 2025
Presents SANDE, a two-stage framework for revoking backdoors in generative large language models while preserving core model capability and avoiding reliance on a clean reference model.
arXiv 2020
Proposes a progressive pose-guided appearance transfer network that preserves garment texture and geometry at up to 1024×1024 resolution without requiring a 3D model.
