Training a Product Library Without a Data Team

Published 2026-07-30 by Faraz Rahimi

You can grow a usable visual catalog from the work people already do, as long as confirmation and rejection write back into the library.

Training a Product Library Without a Data Team

Tags: Computer Vision, Machine Learning, Operations, SKU, unlisted

Nobody is going to photograph 4,000 SKUs on a turntable because you asked nicely. The library has to be a side effect of work.

Every confirmed match can donate a frame. Every reject can donate a "this is not that" pair. Barcode scans can trigger a background photo if the camera is already out.

Start with the ugly SKUs

Do not boil the ocean. Ask leads which items cause radio calls. Those 40 SKUs, with 10 real floor photos each, will move the needle more than 4,000 catalog PNGs.

I keep originals and a normalized crop. The crop trains faster. The original saves you when you change preprocessing later.

Governance is a folder and a rule

A lead should be able to delete a bad photo. If a seasonal hire photographs their shoe, that image should not live forever as "SKU 1182."

Original post: https://farazrahimi.com/posts/training-a-product-library-without-a-data-team