I build open-source tools that make complex systems easier for AI agents—and humans—to inspect.
My current work operates at two scales: software architecture and technical knowledge.
我在构建让 AI Agent 与人类都能看清复杂系统的开源工具：Archiscope 把代码库转化为可校验、可缩放的架构视图，Interlinear 则用行间中文注释、 几何直觉与自适应解释降低技术论文的阅读门槛。共同目标是把隐藏结构变成 可检查、可纠正、可复用的知识。
| Project | What it makes visible | Proof |
|---|---|---|
| Archiscope | The part of a codebase an agent needs right now | 16 terminal-native views · 25 validation rules · Mermaid from one YAML source |
| Interlinear | The terminology hiding inside a technical paper | Reader-adaptive Chinese annotations · geometric intuition · portable agent skill |
AI agents rarely fail because they cannot generate another page of text. They fail because the important structure is invisible:
- dependencies are buried in a repository;
- assumptions are buried in a paper;
- confidence is buried in prose;
- the user cannot inspect what the agent thinks matters.
I am building small, inspectable tools that expose that structure.
codebase ──→ architecture map ──→ focused decision
paper ──→ interlinear notes ──→ continuous understanding
- Architecture intelligence for coding agents
- Portable skills across Codex, Claude Code, and OpenCode
- Human-readable validation instead of opaque agent memory
- Better geometric explanations for technical concepts
Python · Markdown · YAML · Mermaid · GitHub Actions · Agent Skills
- Proof over claims. If a metric cannot be derived from the repository, do not advertise it.
- One source of truth. Views and explanations should trace back to inspectable data.
- Clarity over decoration. A repository should explain itself in under a minute.
- Human override. The user must be able to inspect, correct, and steer the agent.
Explore the repositories · Archiscope issues · Interlinear issues