I build AI systems and developer tools, with a current focus on LLM inference, heterogeneous runtimes, performance engineering, and reliable engineering workflows.
我主要关注大模型推理系统、异构计算后端、性能工程，以及让研发流程更可靠的开发者工具。
Currently: Exploring LLM scheduling, KV-cache correctness, and heterogeneous inference.
- LLM inference & serving — scheduling, KV-cache correctness, distributed execution, and high-performance serving paths in vLLM and InfiniLM.
- AI runtimes from first principles — tensors, operators, C APIs, and Qwen2 inference across CPU, NVIDIA, and CoreX backends.
- Performance & observability — hardware PMU-based runtime telemetry for the JVM and practical performance-analysis workflows.
- Agentic developer tools — repository-aware CI generation and evidence-backed deep research over large code and documentation trees.
| Project | What it does | Stack |
|---|---|---|
| RepoFlow | Detects a repository's language, framework, package manager, and commands, then generates a conservative GitHub Actions workflow. Supports Node.js, Python, and Go projects. | TypeScript, Node.js, GitHub Actions |
| Ancoder Deep Research | Searches code repositories and document trees using dynamic skill generation, multi-round cross-validation, and a test-driven optimization loop. | TypeScript, Node.js, agentic workflows |
| Auto Trader Demo | A research-oriented A-share backtesting toolkit with market constraints, risk controls, parameter search, walk-forward validation, and live screening. | Python, BaoStock, quantitative research |
| LLAISYS | Builds an educational AI runtime from tensors and CPU operators through Qwen2 inference, with NVIDIA and CoreX backends. | C++, Python, CUDA/CoreX, XMake |
- vLLM — fixed DFlash draft KV-cache group annotations and improved executor shutdown behavior during collective RPC.
- InfiniLM — implemented priority-aware request scheduling for the inference engine.
- BiSheng JDK 17 — developed runtime telemetry for Arm PMU events, including optional counters and event-group rotation.
- LLAISYS — implemented tensor operations, CPU operators, Qwen2 inference, and NVIDIA/CoreX runtime backends.
I enjoy working close to the boundary between models and machines: turning scheduling, memory, kernels, and observability into systems that are easier to understand and operate.
Explore my repositories below, or start with RepoFlow for a compact example of how I approach developer experience and reliable automation.