Intelligent product catalog manager

Enjeux

The product catalog lived in two contradictory systems: the professional database on one side, with its negotiated prices, and the online store on the other, with its descriptions and photos. No one owned the full picture, and everyone was right within their own tool. When the online store launched, these discrepancies became visible to customers—even leading to allergens being displayed on the wrong product pages.

Workflow

Single source of truth → Field-level ownership matrix → Decision log → Equipped agents → Dated daily audit

Stack

Ce qu'on a mise en place

Fixing a catalog once is useless. What lasts is a written rule and the audit that verifies it.

A single repository of 628 SKUs. Nothing is deleted: removed lines are deactivated and tagged.
A matrix defines, field by field, which source is authoritative. A dated log records 21 arbitration decisions—without it, agents just invent a plausible rule.
Three agents with specific write permissions: one who responds, one who proposes and only writes after validation, and one who audits but never corrects.
Write access lives in the connection, not the prompt. A read-only token cannot write; an instruction is merely an intention.
All bulk writes are verified by recount. The rule was born from an incident, detected by a counter that didn't add up.

Impact

A single source of truth is authoritative. Zero duplicate product codes, zero broken links out of 358, zero divergent compositions out of 356.
No more internal labels visible to the customer. There used to be 30.
The state of the catalog is captured in a dated report every morning. Previously, it wasn't measurable.
242 product pages are awaiting human input. The system doesn't correct them: it identifies them and prevents new ones from being added.

Have a project, a question, or just want to bounce some ideas around?