Student researcher and builder working on AI agents, game technology, and developer tools.

I'm an undergraduate in Artificial Intelligence at MBZUAI, based between Abu Dhabi and Beijing. I build across agent systems, games, robotics, and full-stack software. I like projects that begin as uncertain technical questions and end as something people can inspect, test, or play. I document what worked, what failed, and what remains unproven.

Website · Research · Projects · Notes · CV · LinkedIn · Email

我喜欢把不确定的技术想法变成可以验证、可以解释，也可以真正使用的系统。

• Hardware agent harnesses. How can natural-language tasks become constrained, testable, and auditable robot skills that move from simulation through safety gates to real hardware?
• Adapting motion foundation models. Which data, training, and evaluation loops can extend KIMODO toward martial arts while exposing failure as clearly as success?
• Agent-native organizations. If autonomous agents become economic participants, how might labor, ownership, productivity, and the distribution of wealth change?

These are active research questions, not claims of completed systems.

• ModCraft — A product for making game mod creation more approachable. I work as CPO and primary full-stack engineer across product, engineering, QA, and DevOps.
• OpenDoc — A team-built real-time collaborative editor. I owned backend architecture, authentication, permissions, sharing, collaboration infrastructure, and developer tooling.
• Arcane Bond — A local two-player pixel action game built around asymmetric spell-casting and shared-screen play.
• CreditSense — A reproducible tabular-ML pipeline for jointly predicting credit-risk class and loan interest rate, with cross-validation, feature engineering, and model tuning.
• Twelve Bars — A browser-based rhythm section for jazz and blues practice, built with React and the Web Audio API.

More public builds are collected on my projects page.

I publish scoped case studies with the experiment conditions, evidence, limitations, and negative results left visible. Most current notes are in Chinese.

• UE 5.8 official MCP: from a large tool surface to verifiable workflows
• AnimGen: from animation-data filtering to runtime pose generation
• Robust Motion In-Between: from a reproducible benchmark to UE 5.8
• KIMODO capability evaluation: prompts, constraints, and boundaries
• A practical map of AI animation: generation, in-betweening, retrieval, constraints, and runtime control

The full collection lives in the AI animation research archive.

• Agents and backend: Python, FastAPI, LangChain/LangGraph, MCP, PostgreSQL, Redis, Docker
• Product and web: TypeScript, React, WebSockets, Yjs
• Game and ML research: Unreal Engine, PyTorch, ONNX, NNE

Source first, then small verified loops. I care about prototypes that explain the idea, expose their limits, and remain useful after the demo.

Outside technical work, I keep returning to music and improvisation, philosophy, and history.

I'm currently open to research collaborations at MBZUAI and AI systems internships for Summer 2027.