I build AI agent systems for real-world work.
evidence → judgment → action → result
I am interested in the part that begins after a model gives a good answer: connecting reasoning to tools, state, permissions, people, and outcomes.
My work moves across commerce, research, creative production, knowledge workflows, and software delivery. Different domains, same question:
How do we turn intelligence into dependable action?
- Evidence before confidence. Important decisions should be explainable.
- State before magic. Work should survive pauses, failures, and handoffs.
- Boundaries before autonomy. Agents need clear tools, permissions, and stopping points.
- History before overwrite. Inputs, decisions, edits, and results should leave a trail.
- Reality before polish. A small system that works in practice beats a perfect-looking demo.
- Building reusable foundations for stateful, tool-using agents
- Testing them inside real operating workflows
- Open-sourcing selected building blocks as they become ready
- Writing down the choices, failures, and lessons behind the systems
The interesting part of an agent is not just what it can say. It is what it can carry forward, what it is allowed to change, how it recovers, and whether someone can understand what happened afterward.
That is the work I am here to explore.