I build reliable AI agents and the tools around them.

I build small, open-source tools for the parts of agent engineering that are easy to ignore and expensive to get wrong: context, evaluation, safety, and developer experience.

A control plane for the tools agents run — inventory, evidence, policy and enforcement — together with a daily-rebuilt evidence index of the public MCP ecosystem.

Everything in it is organized around one rule: clean is only emitted when every check ran. A record that could not be measured says what stopped it, and a repository record can never be clean, because nothing read its source. Zero runtime dependencies, and the whole test suite runs from a clone with nothing installed.

A local, zero-API-key audit for agent instructions, skills, and MCP configuration, including a bounded model of the Codex instruction chain.

It estimates instruction-context cost, catches duplicated instructions, flags risky configuration, and produces a scorecard that can run in CI.

I built a small set of independent, open-source tools for the DeepSeek Harness ecosystem:

• dsh-doctor — local environment, profile, and session-log diagnostics; it also helped align the shared dsh-doctor/v1 community envelope.
• dsh-trajectory — portable, self-contained HTML trajectories from Harness session logs.
• dsh-report-studio — verifiable daily, weekly, handoff, and article reports from session evidence.
• dsh-plugin-starter — a zero-dependency scaffold that encodes common plugin-development pitfalls.

The public record includes a plugin field guide and the dsh-doctor interoperability discussion, a multi-frame session-log format proposal, and merged community-directory contributions. These are independent, unofficial community projects, not official DeepSeek releases.

• Ship runnable tools, not AI theater.
• Never report clean about work that did not run; a gap has to name itself.
• Measure context before adding more context.
• Keep the default path local and inspectable.
• Treat evals and failure cases as product features.
• Support the agent stack people already use.

If you are working on agent infrastructure, open an issue or start a discussion.