Incident-AI: the AI proposes, a human approves
Alert in → persist → proposal → human gate → traces. Drafts start unsent. Close needs a person.
Watch the walkthrough (~14 min)
Read the essay on Medium: “AI proposes. A human authorizes.”
Short clip: model down, ticket survives (~3 min)
See it handle this site's real incidents, live →
Postmortem: game day 1, where I broke the site on purpose and it raised and triaged the incident.
The model only proposes. Since 1 Oct 2026, every AI priority on the live desk waits for a human approval (the one earlier ticket is labelled with the rule it ran under), and AI priority never pages anyone: paging comes from an outside probe on this site's SLO.
Code: vikiscoding/vikrant_perswebsite_incidents_aiengine
Scope: a Python engine running one site's incidents, not ServiceNow.