I make software safe to run when you can't fully trust what it will do.

Selected work

1. Bounded Systems

   Capability security for coding agents

   Problem

   Agent-authored commits reached main without review, and afterwards nobody could say which writes the agent had been authorized to make.

   Intervention

   Moved privileged git effects behind capability-checked operations, and routed the work through one pipeline whose only exit is a merged pull request.

   Evidence

   prx drives a work unit to an attributed pull request; ocap-provenance records which capabilities each effect held, and which were denied. Both are public, and both name what is still outside the check.

   • prx
   • ocap-provenance

   Role

   System model, schemas, implementation, and the operational testing.

2. Commerce systems

   Supply planning and inbound logistics

   Problem

   A 9-tab spreadsheet was the supply plan of record, and three teams each read it differently.

   Intervention

   Replaced it with a CLI over one source of truth — 61 schema contracts, 173 test scenarios — and automated the inbound freight it had been gating.

   Evidence

   The three teams now plan against the same document, and freight moves without the manual coordination that used to hold shipments.

   Role

   Contract design, the CLI, and the cross-team consolidation.

3. Developer systems

   Quality gates for a 20-engineer team

   Problem

   Whole classes of defect were being found after merge, by people, one ticket at a time.

   Intervention

   Moved static analysis into the pull-request pipeline as a blocking gate, so those classes could not reach main in the first place.

   Evidence

   Critical bugs fell 20% — critical-severity tickets per quarter, the four quarters before against the four after.

   Role

   Gate design, the CI implementation, and the developer-experience work that kept the gate from being routed around.