I build AI agent systems.

I’m a Senior Staff Engineer at Alibaba. Our team works on two fronts: lifting the limits of what LLMs can do through new applications and agent harness design, and adapting agents to specialized domains through training, evaluation, and secure runtimes. Current projects include QwenPaw, an open-source, secure, self-improving personal assistant, extensible through a large ecosystem of plugins; and TuFT, a multi-tenant platform for fine-tuning LLMs through a unified API. Our work is open source under the AgentScope organization.

Previously, I was a Principal Researcher at Microsoft Research in Redmond, Washington, where I served as the lead architect of AutoGen, an open-source framework for building AI agents and multi-agent applications. AutoGen offers a high-level API for orchestrating agents that collaborate autonomously or with human supervision, an event-driven, low-level API for full control over agent behavior, and a distributed runtime, built on the Actor Model, that hosts agents written in different programming languages. AutoGen grew into a widely adopted tool for agent developers and researchers, shaped by a community of over 500 contributors.
AutoGen:

I also worked with Azure AI to build the Foundry Agent Platform, a cloud platform for deploying AI agents and workflows at scale. As part of this effort, I helped create the Microsoft Agent Framework, an evolution of AutoGen for enterprise-grade applications, which now sits at the platform’s core.
Agent Framework:

Earlier at Microsoft Research, I worked on database systems. My collaborators and I developed a cost-based, platform-independent rewrite rule for MATCH_RECOGNIZE queries in general-purpose SQL engines, achieving a 5.4X median latency improvement in Trino. We then built a specialized execution engine for MATCH_RECOGNIZE with extended operators and a novel cost-model-based optimizer, delivering a 6X median performance gain over state-of-the-art engines.

Before joining Microsoft Research, I completed my PhD in Computer Science at the University of Toronto, advised by Prof. Renée J. Miller. My thesis focused on dataset search over massive Open Data archives. I contributed algorithms for large-scale set similarity search and data sketches, capable of searching over 100K tables for joinable or unionable ones in milliseconds. I also built an Open Data search engine stack to make Open Data more accessible for downstream applications.
SetSimilaritySearch:
datasketch:

I build in the open. Through open-source projects and writing, I aim to make advanced AI and algorithms accessible to everyone.

[GitHub] [Google Scholar] [Blog] [Email]