About

A mathematical route into systems

My doctoral research at Zhejiang University focuses on high-order numerical methods for partial differential equations on complex geometries. Building those solvers meant going beyond derivations: I designed C++ abstractions for geometry and operators, debugged sparse linear systems, parallelized kernels with OpenMP, and built convergence and performance experiments that could survive repeated refinement.

That work shaped how I approach engineering problems. I first make the computational model explicit, then identify correctness boundaries, measure the real execution path, and optimize only where the data points.

From numerical solvers to distributed execution

During my internship at Tencent's Distributed Database R&D Center, I applied the same approach to a production MPP analytical engine. My work covered distributed aggregation and complex-plan adaptation, query profiling across execution layers, and compute-node reliability. The domain changed; the central questions did not: where is state owned, which work can run independently, how is correctness preserved across boundaries, and what does the profiler actually prove?

Current direction

I am targeting AI infrastructure, performance-critical C++ systems, and quantitative-development roles. My current study extends from PyTorch and deep-learning fundamentals toward LLM inference—prefill/decode behavior, KV-cache management, prefix caching, serving schedulers, and the vLLM/SGLang codebases.

How I work

Skills in context