Work that ships.

A language model can reason about a screen but can't physically operate one. PiCUA closes that gap: an AI agent that drives real machines through PiKVM hardware — so a model can boot, control, and recover a computer at the firmware level, not just click around inside an already-running OS.

I lead the AI infrastructure and the computer-use agent itself, with a bias toward real-world deployability over demos — intelligent automation that lets an agent interact with a physical machine remotely. The result is the world's first physical computer-use AI agent, live at picua.ai.

An agent that can use tools is also an autonomous attack surface. I lead red-team and security research across agentic AI systems and computer-use agents — probing where a model can be steered into doing something it shouldn't.

The work turns offensive findings into concrete defenses: hardening autonomous agent pipelines against prompt-injection and tool-abuse vectors before they reach production.

Responsible AI needs humans in the loop where it matters. I built human-feedback-driven systems and the analytical admin tooling that gives teams oversight of AI workflows — alongside the evaluation frameworks and quality testing that keep model behavior honest.

The implementation went on to win a 2026 Webby Award (People's Voice) for Best Responsible AI.

Agentic commerce needs a protocol agents can actually transact over. I led engineering for EasyACP — built on top of the Agentic Commerce Protocol — coordinating the team and the backend architecture behind AI-assisted commerce workflows.