My own operations run on an AI operating system I built: more than 100 scheduled jobs behind a control plane with spend caps and a kill switch, and typed approval before any agent-proposed change merges.
Scheduled jobs, scripts and AI agents run my operations. Every one of them is registered in a control plane with spend caps, a kill switch and typed approvals.
When an agent proposes a change, it arrives as a review card. It merges only after my typed approval, and it is measured afterwards.
Since 28 September 2026, a weekly scorecard has audited my AI operator's replies: corrections per 100 replies by model, completion claims checked against evidence, golden questions, and missed replies.
Its first study tested the model-switch theory and found no cluster of corrections after a switch.
The top correction type from the scorecard becomes one proposed rule change.
A replay gate tests the change on golden questions and past incidents.
The change merges only on my approval.
Built on 4 October 2026. The first card was wrongly denied by an absolute bar, so the gate now judges against the baseline.
I work US hours and I am open to AI automation engineer roles and client projects with international teams.
Work with me Back to portfolio