I'm a software engineer interested in coding agents, agent evaluation, tool use, and reproducible AI systems.
My current work focuses on backend engineering and AI-assisted systems. Outside work, I build open-source tools to explore practical questions about coding agents:
- Do skills and prompts actually improve task outcomes?
- When does an MCP server help, and when does it only add latency?
- How can coding-agent experiments be reproduced across environments?
- How should we evaluate agent reliability, cost, and effectiveness together?
I'm building AgentAblate, an open-source framework for reproducible ablation testing of coding-agent skills, prompts, tools, and MCP servers.
It helps developers compare agent configurations while tracking outcomes, runtime behavior, cost, and reproducibility.
- Coding agents and AI-assisted software engineering
- Agent evaluation and benchmarking
- Tool use and Model Context Protocol (MCP)
- Reproducible AI and empirical software engineering
- LLM reliability, efficiency, and human-agent collaboration
Central South University · 2019–2023
- GPA: 85.23/100
- Huawei Intelligent Foundation Scholarship
- Outstanding Student Leader, Central South University
JD.com · July 2023–Present
- Develop and maintain backend systems for financial risk-control services.
- Work on the engineering and application of JD AI Coach.
- Explore how AI capabilities can be integrated into practical business workflows.
- Focus on system reliability, maintainability, and production-oriented delivery.
Meituan — Food Delivery Technology Department · May 2022–November 2022
- Worked on the store operations module of the Meituan Food Delivery merchant platform.
- Participated in the development and maintenance of merchant-facing backend services.
- Collaborated with product managers, frontend engineers, and backend engineers to deliver features for restaurant store operations.
- Gained practical experience with production development, testing, code review, and release workflows.
A local experiment harness built around one practical question:
Did this skill, prompt, MCP server, or tool policy actually improve the coding agent?
Current capabilities include:
- Reproducible baseline and treatment runs
- Coding-agent adapter support, including Codex
- Runtime identity and environment fingerprints
- Skill and MCP configuration validation
- Structured experiment reports
Contributions, experiment ideas, and feedback are welcome.
- Backend engineering and distributed systems
- Python and Java
- Coding agents and tool-using AI systems
- Model Context Protocol
- Experiment design and ablation studies
- Testing, automation, and reproducible engineering