Not one person —
a team of AI agents.

PLATFORM WORK

Enterprise agentic AI platform

Making agents orchestratable, observable, and evaluable inside real companies — not chat toys, but digital workers that ship to production, submit to governance, and account for their cost.

• Multi-agent orchestration

  Task decomposition, role assignment, parallel execution, and clear boundaries for human-in-the-loop control.

• Observability

  Every decision and tool call traceable, so failures resolve to root causes instead of vibes.

• Evaluation

  Reproducible evals in place of "it feels better" — iteration backed by evidence.

One operating principle: manage AI as staff, not as tools. My agents have roles, discipline, and performance reviews — coding, reviewing, and operating in separate seats — while I do governance and final calls.

SELECTED WRITING

Methods and failures, in public

• Code Is Cheap, Governance Is Not [zh]

  Where engineering value moves once AI floors the cost of writing code.

  ->

• What Is a Token? The Credential of the AI Era [zh]

  Reasoning backwards from the billing unit to model capability.

  ->

• AI Sycophancy: More Dangerous Than the 80% Trap [zh]

  The moment the model agrees with you is the moment to be careful.

  ->

• If You Still Have to Remind People, the Discipline Is at the Wrong Layer [zh]

  Put discipline into the system, not into memory.

  ->