Train a product categorization model for a 180k-SKU marketplace catalog

Our marketplace aggregates home improvement products from about 60 suppliers, and every supplier categorizes things differently - one vendor's "fixtures" is another's "bath hardware". We have 180k SKUs, roughly 40k of them hand-labeled into our 900-node category tree by our ops team over the past year, and we want a model that classifies the rest and handles new inbound feeds. The work: explore the labeled data, pick an approach (we suspect fine-tuning a small transformer on title + description + supplier attributes will beat classic TF-IDF, but convince us with numbers), train it, and hand over a documented Python inference service plus a confidence-thresholded review queue design so low-certainty items route to humans. We measure success as >=92% top-1 accuracy on a held-out set our team curates. Expect an async-first engagement. We will give you a sanitized data export on day one and access to a staging database. Weekly written progress updates matter more to us than meetings.