BUILD · EVALUATE · SHIP

Engineering Experience

I build AI systems that can be evaluated, operated, and improved: coding-agent benchmarks, repository understanding pipelines, retrieval systems, and production full-stack applications.

This site focuses on engineering work and implementation evidence. Publications, academic CV, research interests, and scholarly updates live on my academic profile.

Academic profile: publications, CV & research ↗

Engineering Experience

2024 — 2026
2026.03 — 2026.06
Beijing, China

ByteDance · Seed

Algorithm Engineering Intern · Code Evaluation

  • Built task-generation and evaluation pipelines for feature implementation, test generation, and repository-level code understanding, including automated environments, multilingual support, and distributed execution.
  • Designed a four-stage rule-based difficulty funnel for millions of test-generation candidates. The filtered score distribution matched Best-of-N while substantially reducing its inference dependency.
  • Developed a repository-understanding pipeline that parses documentation and code structure, locates key modules, and produces verifiable facts and QA pairs for tens of thousands of mid-training and RL samples.
  • Led a frontend Agent-as-a-Judge pipeline that converts prompts and generated pages into UI tasks and rubrics. A 2K+ black-box benchmark achieved 90%+ human–AI scoring agreement.
2025.07 — 2025.10
Montreal, Canada

REALISE Lab · Concordia University

LLM Systems Research Intern

  • Built a retrieval system with 5,000+ instruction–code pairs and API documents to address long-tail knowledge gaps in low-resource library code generation.
  • Designed a Retriever–Generator multi-agent workflow for automated prompt orchestration, reducing code-generation errors by up to 25% across libraries including Plotly.
2024.06 — 2025.07
Beijing, China

HCI & IIP Lab · ISCAS

LLM Application Engineer · Research Assistant

  • Led the interaction module for an LLM-enhanced low-code agent platform, implementing multi-turn conversation and memory with LangChain and integrating it with the wider system.
  • Built vertical instruction-tuning data and applied LoRA to Llama 3–8B, improving instruction-following accuracy by 10%. Implemented a Neo4j knowledge graph and error-attribution pipeline spanning 100+ concepts.

How I Build

PRINCIPLES
01

Evaluation first

Define executable checks, benchmark design, and failure modes before optimizing model behavior.

02

Systems over demos

Connect data production, model calls, orchestration, observability, and interfaces into repeatable pipelines.

03

Measure the bottleneck

Use profiling, controlled experiments, and quality metrics to decide what to improve next.

Technical Skills

STACK

Languages

PythonC/C++JavaScript / TypeScriptJavaSQL

AI Systems

PyTorchLangChainTransformersRAGLLM AgentsLoRAEvaluation

Production

ReactFastAPINode.jsMySQLDockerNginxCI/CDGit