The enrichment machine.
Half a million SKUs, decades of accumulated catalog, and thousands of products that exist as little more than a part number and a price. You can’t sell what you can’t describe — and no team on earth fixes that by hand. So we built a machine for it.
The architecture is a pipeline with a confidence gate. Every sparse record goes first to a web-research model, which hunts down what the product actually is — manufacturer data, specifications, descriptions — and drafts a complete record with sources. If the result comes back high-confidence, the record completes automatically. No human touches it.
Low or medium confidence? The record escalates to a second, heavier model for a deep-research pass. High confidence there → complete, again untouched. Only the records that neither model can settle land in a human review queue.
The part worth underlining: we didn’t assume the automatic path could be trusted — we tested it until it proved itself. That’s the same standard I hold AI to everywhere: it earns its place by being verified, not assumed. The models do the volume. Confidence decides the routing. People spend their judgment only where judgment is genuinely needed.
Next entry: teaching AI to read forty years of printed catalog pages.