I build AI agents that ship — enterprise workflows in production, and open-source agent tooling. 设计并交付真正上线的 AI Agent 系统——从生产环境的企业工作流，到开源 agent 工具。
By day, I lead enterprise AI solutions from PoC to production at a Fortune Global 500 technology company — as solution architect, core developer, and delivery owner. Before that, I was a product manager for overseas consumer AI products (AI chat, image generation and more) — one made the a16z global Top 50, another reached six-figure monthly revenue. I build mainly in Python and TypeScript, and work daily with AI coding agents such as Claude Code and Codex — the open-source products below grew out of my own agent workflows. 现职于一家世界 500 强科技公司，负责企业级 AI 方案从 PoC 到生产上线——身兼方案架构、核心开发与推进 owner。此前是海外 AI 消费产品的产品经理（AI 聊天、图像生成等），其中一款产品进入 a16z 全球 Top50，另一款月收入达六位数美元。主要使用 Python 和 TypeScript，日常深度使用 Claude Code / Codex 等 AI coding 工具——下面的开源产品正是从我自己的 agent 工作流中长出来的。
01 · Products01 · 在做的产品
Tooling for the agent era, extracted from my own workflows.为 agent 时代做的工具，全部来自我自己的真实工作流。
Context engineering for coding agentsCoding agent 的 Context 工程
A portable role-context protocol. Define your product, roles, domain knowledge and decisions once — compile them into context packs for Claude Code, Codex and other agents, instead of re-explaining everything every session. 可移植的角色上下文协议。把产品、角色、领域知识与决策一次性结构化，编译成 Claude Code / Codex 等工具的 context pack——不再每个 session 从头解释背景。
$ npm install -g contextspecLong-term memory & triggers for agentsAgent 的长期记忆与事件触发
Long-term project memory for coding agents. One markdown file per project plus a global watchlist with date and event triggers — your agent syncs everything you're pursuing across folders and months. coding agent 的长期项目记忆。每个项目一个 markdown 文件，加一张带日期/事件触发器的全局 watchlist——让 agent 跨文件夹、跨月份同步你所有在推进的事。
02 · Open source02 · 开源贡献
Contributor to the following agent-tooling projects.已在以下 agent 工具项目中进行贡献。
An agent-harness performance-optimization system for Claude Code, Codex, Opencode and Cursor — agent skills, learned instincts, long-term memory, security gates, and research-first development workflows. 面向 Claude Code / Codex / Opencode / Cursor 的 agent harness 性能优化系统——agent skills、习得式 instincts、长期记忆、安全门控与 research-first 开发工作流。 3 merged contributions · agent tooling & runtime3 项已合并贡献 · agent 工具链与运行时
A multi-agent framework for agile AI-driven development — role-based agents (analyst, PM, architect, developer) and structured workflows from planning and specs through implementation. 敏捷 AI 驱动开发的 multi-agent 框架——角色化 agent（分析师 / PM / 架构师 / 开发者）与规划到实现的结构化 workflow。 Merged contribution · agent config resolution已合并贡献 · agent 配置解析
03 · Selected work03 · 工作案例
Lead architect · core developer主架构师 · 核心开发
A root-cause analysis system over customer-feedback text: layered multi-dimension LLM pipeline combining rule logic, model reasoning and business heuristics; prompt engineering with a systematic evaluation-and-tuning loop; fully automated end-to-end runs. Accuracy up 20+ percentage points over the single-pass baseline; tens of thousands of records processed. 客户反馈文本根因分析系统：多维度分层 LLM pipeline，规则逻辑、模型推理与业务启发式结合；prompt 工程配套系统性评测与调优闭环；端到端全自动运行。准确率较单次 prompt 基线提升 20+ 个百分点，已处理数万条记录。
Solution architecture · core implementation方案架构 · 核心实现
A LangGraph-based multi-agent system, developed skill-driven: file ingestion, backend and API integration, structured data processing, and result presentation. 基于 LangGraph 的 multi-agent 系统，采用 skill-driven 开发——文件接入、后端与 API 集成、结构化数据处理、结果展示。
End-to-end owner端到端负责人
Ran an internal AI hackathon end to end with dozens of teams, delivered enablement sessions reaching hundreds of attendees, and supported day-to-day AI-tool onboarding and consultation across the organization. 端到端负责数十支队伍参与的内部 AI Hackathon；组织覆盖数百人次的 enablement 分享，支持组织内日常的 AI 工具 onboarding 与咨询。