I build languages, compilers, and runtimes for agentic systems. My work focuses on reliable agent execution, skill compilation, runtime evaluation, and efficient inference on real hardware.
- Academia — Academic M.S. in Computer Science and Technology, Sun Yat-sen University · 2027 Entry. Research with Prof. Xianwei Zhang at arcSYSu Lab, National Supercomputer Center in Guangzhou.
- Industry — AI Agent R&D Intern, Band AI, ByteDance · Jul–Nov 2026.
- Open source — Building Nexa and its agent language, compiler, and runtime ecosystem.
| Project | Focus |
|---|---|
| Nexa | A harness-native language for agent workflows, with compiler checks, a Python runtime, and a Rust Agent Virtual Machine. Docs |
| SLIM-ARC | Memory-constrained MoE inference using OS virtual memory and sparse expert loading. National First Prize, 2026 Operating System Design Contest. |
| ASCEND-OK | Evolutionary kernel optimization for Triton-Ascend, with compilation, correctness checks, and measurements on NPUs. National Third Prize, 2026 Compiler System Design Contest. |
| ClassFox | A local-first desktop classroom assistant for transcription, contextual Q&A, and Markdown notes. Website |
| YatCC-AI | Compiler teaching and LLM/HPC tools. Leading the AI Hub gateway and teaching documentation. |
- SkCC — Portable skill compilation across agent frameworks. Accepted at ACM CAIS 2026 — AgentSkills'26 Workshop. Project
- RAMP — Runtime assessment of agentic models in production systems. Submitted to HPC Asia 2027. Project
- Also working on agent memory and context management, topology-aware skill routing, and GPU kernel fusion. More on my homepage.
Tools I work with: Python · C/C++ · Rust · CUDA / Triton · LLVM · Linux
A few graphs, pixel scenes, and contribution games — just for fun.
Streaks cover the past year in UTC; the activity graph covers 31 days. The summary card includes all-time history. Working hours use UTC+8. Totals can differ by generator and GitHub view.
Outside coding: calligraphy, badminton, and geography. Support my work.