AI engineer building reliable agent systems and production-grade developer tools.
10 years in software engineering · Architecture · Evaluation · Reliability
I work where AI prototypes become dependable systems: agent architecture, evaluation, orchestration, and the controls that make long-running autonomous work inspectable and verifiable.
我做可评估、可追踪、可交付的智能体工程，重点是长任务里的 context death 和验收收敛。
longgraph-skill — long-horizon agent skill for Claude Code, Cursor, Codex, and Grok Build. Clean-context supervisor, multi-task ledger loop, verified gates. A markdown library (loop-graph), not a framework.
| Project | What it does |
|---|---|
| longgraph-skill | Long-horizon agents: clean-context roles, durable state, /loop-converge for code-convergence. |
| skillmill | Docs site, PDF, or tutorial → installable agent skill (npx skillmill). |
| session-skill | Successful agent session (chat + tool traces) → installable SKILL.md for Cursor / Claude Code. |
| agent-ding | Notify when a coding agent finishes — modular hooks, Zellij, shell helpers. |
| dsh-plugin-longgraph | DeepSeek Harness plugin: longgraph / loop-graph / loop-converge authoring skills. |
| Project | What it explores |
|---|---|
| obsidian-llm-wiki | Maintainable knowledge system: namespaces, layered indexes, reusable agent skills. |
| sherlock-claude | Repo + runtime logs → focused diagnoses and fix recommendations. |
- Long-horizon agents — survive context death, keep state durable, make “done” checkable
- Loop converge — bind each loop to a real acceptance gate, not empty code churn
- Skill tooling — turn docs and good sessions into installable skills
- Evaluation & reliability — evidence-based acceptance and failure analysis
Make the state inspectable. Make “done” verifiable. Keep the system smaller than the problem it solves.