We worked with a marketplace with 3 million monthly users.

The problem

Sellers chose categories by hand, and a third of new listings landed in the wrong place. Search quality suffered and moderators spent their days moving ads around.

What we built

A fine-tuned multilingual classifier trained on six years of listings, with synthetic examples for the 300 rarest categories. It runs on CPUs behind a simple API, so the client’s platform team could deploy it without new infrastructure.

The result

Category accuracy went from 71% to 96%. Moderation time dropped by two thirds, and the client’s own engineers now retrain the model every quarter.