I build the reliability layer around coding agents: tools that help them understand code, cross runtime boundaries, act with guardrails, and leave behind work that people can inspect. I'm now turning that systems mindset toward physical intelligence—systems that perceive, form models, take action, and learn from consequences in the world.
My current work follows one continuous path:
source-backed context → bounded action → observable feedback → grounded learning
I'm less interested in making an agent sound capable than in making its claims and actions easy to verify.
Physical intelligence is the next loop I want to close: perception → world model → policy → action → feedback.
Recent curriculum work has taken me through GridWorld, GAE/PPO, continuous-control mechanisms such as DDPG, TD3, and SAC, and model-based ideas including World Models and MuZero. The unfinished parts are exactly what pull me forward: moving from diagrams and simulated loops toward systems that can sense and act in the physical world—without giving up observability, safety boundaries, or evidence.
| Project | The question behind it |
|---|---|
| AgentSession | What actually happened during an agent run? Built on OpenSession, it is a local-first, multi-provider archive for searching, tracing, and replaying coding sessions—including tools, prompts, tasks, and nested agents. |
| CodeFacts | What does the source code prove? Compact, attributable structural facts for coding agents, exposed through a deliberately small read-only MCP surface. |
| Pi WSL MCP | How can Windows-hosted agents work with Pi in WSL without flattening its runtime? A bridge that preserves Pi's sessions, models, extensions, credentials, and read/write boundaries. |
| pi-access-guard | How can a local agent stay useful without getting unrestricted access? An allow / ask / deny guardrail for filesystem and shell operations. |
| pi-weixin-cli | How can an agent meet the workflow where it already happens? A bidirectional bridge between WeChat messages and persistent Pi sessions. |
- Evidence before confidence. Source-backed facts, real protocol smokes, and live UI checks beat plausible prose.
- Local-first by default. Reuse the sessions, models, extensions, and credentials already present instead of copying or obscuring them.
- Small surfaces, sharp boundaries. Fewer tools with explicit read-only and writable paths are easier for both people and agents to understand.
- Build from friction. When a workflow problem repeats, I turn it into a bridge, inspector, or guardrail that can be tested and reused.
Working across TypeScript, Node.js, MCP, SQLite, Python/PyTorch, Windows, and WSL.