AI Agent & LLM Engineering
Building reliable AI applications and the systems behind them.
My work spans AI coding tools, LLM applications, and backend engineering. My background includes cloud / IaaS and GPU-related development at Baidu, and AI coding tools at Alibaba Taotian.
Focus: Agent workflows · AI Code Review · SDK reliability · CI/CD · Retrieval & memory
A selection of merged pull requests, focused on correctness, compatibility, and regression prevention.
| Project | My contribution | Merged PRs |
|---|---|---|
| AMD GAIA | Fixed a release-test gate that silently swallowed failures; corrected backend-test dependencies and added workflow contract tests. | #3573 |
| PR-Agent | Fixed inline-comment line resolution in the Bitbucket Cloud and Server providers, preserving the existing integer-line path. | #3136 |
| OpenSandbox | Bounded readiness-polling sleeps by the remaining timeout budget in the Go, JavaScript, and Kotlin SDKs; added regression coverage. | Go #1744 · JS #1745 · Kotlin #1747 |
| Semantica | Removed redundant vector deletion during memory erasure; removed signed Neptune authentication material from logs and added regression tests. | Memory #1379 · Security #1373 |
| Apache Superset | Isolated table-cell formatting errors so a problematic value can fall back safely without disrupting other cells. | #39658 |
Explore more merged contributions →
Workspace templates for Claude Code and Codex CLI, with separate layouts for agent instructions, MCP setup, and shell-driven coding loops.
The repository keeps the two toolchains separate: cc/ for Claude Code and codex/ for Codex CLI.
Reproduce the failure → understand the contract → make a focused change → add regression coverage.
I care about explicit failure handling, predictable timeout behavior, backward compatibility, and tests that catch the original bug—not just exercise the happy path.
Working stack: Python · Go · Java · TypeScript / JavaScript · GitHub Actions
Open to technical discussions about agent tooling, AI-assisted development, and reliable LLM systems.