I helped design the triage loop behind a platform desk that now carries 12 lanes of work: data model, storefront feed, search, pricing, ops tooling. Requests are classified, routed to the right codebase, run through an agent, and stopped at a human review gate with a rollback snapshot. The agents operate it; the gate is human. Median cycle time 20.7 hours, and what used to live in inboxes now has state and an audit trail.

No single sellable signal existed across the PIM, the feed, and the storefront, and the layers disagreed. Mapped where they diverged, wrote the one-way rule (a child SKU can be discontinued under a live parent, never the reverse), retired 175 SKUs against it without deleting a record, then re-audited the live storefront for leakage: discontinued items still buyable. Zero left.

Took search from patched synonyms and redirects to a rebuilt system. Relevance, suggestion quality, and flagship visibility improved on every metric we tested, and the regression class that buried products under blog noise stopped reproducing. One query that returned 337 suggestions now returns 16.