I lead the effort to build the agent foundation from the ground up at a post-transformer AI lab. My current research is on long-running agents, multi-agent systems, agent harnesses, and self-improving AI systems.
I work with academic institutions on interdisciplinary AI research, including AI for science with the Shenzhen Institutes of Advanced Technology (SIAT).
Previously, I founded Miototech in China, one of the earliest AIGC companies, backed by Baidu VC, and Currents AI in the US, backed by VCs and part of NVIDIA Inception. Earlier in my career, I was an engineer at TikTok and AWS, where I built large-scale production systems from the ground up. At TikTok, I built core infrastructure from 0 to 1 that grew to support over 100 million daily active users; at AWS, I built services serving customers across 22 countries.
I am a cat dad to ChouChou
and Haru
, and a beginner tennis player.
Research
Self-Improving Harness
I build self-improving harnesses that close the loop between agent execution and learning. The harness turns runtime traces into structured feedback, evaluation, and iteration, so experience accumulates across tasks as durable capability gains rather than one-off completions.
AI for Science
In collaboration with the Shenzhen Institutes of Advanced Technology (SIAT), Chinese Academy of Sciences, I work on AI for science. I build multi-agent systems that automate the research loop — running experiments, checking results, and iterating. neuroasd is one example, applied to autism brain-network analysis. This research is supported by the Natural Science Fund of Guangdong Province.
AI-Native Infrastructure
I investigate infrastructure designed natively for AI agents rather than adapted from traditional web systems. This includes agent-oriented interfaces, structured interaction protocols that significantly reduce token consumption, and storage architectures integrating embeddings, graphs, and persistent memory to support scalable agent computation.
Long-Horizon Reasoning Systems
For agent systems that have to keep working across long tasks, I look at how context is assembled, how retrieved graph structure is used during inference, and how memory stays consistent as work proceeds. Multi-agent handoff is part of that, so state does not collapse back into a single prompt.
Reliable AI Applications
I examine how to translate foundation models from benchmark settings into dependable production systems. This involves improving robustness, controllability, and observability, and grounding generation in structured signals to ensure reliability in real-world deployment.
Systems Built
- Currents Social Intelligence Engine (currents.one). Built an end-to-end social intelligence system for large-scale discussion analysis, integrating hierarchical long-context modeling, graph-aware retrieval, and emerging segment detection. Extracts structured product signals and user clusters from noisy social data. Launched on Product Hunt and ranked #1 of the day.
- Web Agent Harness. Built a web agent harness with GPT-5 that achieved state-of-the-art results. Proposed and implemented treating the browser as a code execution environment for agents, rather than simulating human UI actions — predating similar approaches later popularized by systems such as Browser Use.
- AWS SageMaker No-Code ML Platform. Built a no-code machine learning platform on Amazon SageMaker, enabling users to train, evaluate, and deploy models through a visual workflow instead of writing code.
- TikTok Live Stream Architecture. Early member of the TikTok Live engineering team. Led the 0 to 1 design and implementation of the core user authentication system for live streaming, supporting infrastructure at billion-scale daily active users.