5,977 products matched to their photos
A UAE minimart had 5,977 products in its point of sale and about 33,700 product photos with random file names and nothing to join them on. Our AI shortlisted likely photos for every product, a vision model picked the right one with a confidence score, and uncertain matches went to a person.
- Client
- A UAE minimart
- Industry
- Retail & e-commerce
- Timeline
- One overnight run, after a 200-product pilot
The situation
The shop was moving its catalogue online and needed a photo on every product. The photos existed, more than 33,000 of them, but their file names were random and there was no barcode or code that linked a photo to a product.
What was broken
- No shared code between the product list and the photo files.
- Tens of thousands of images to search for each of almost six thousand products.
- Products had no department or category, so the online shop could not be browsed.
What we built
Visual search shortlist
Every photo was turned into a searchable fingerprint, so each product name returned its 20 most likely photos instead of a search through the whole folder.
A vision check with a confidence score
A vision model looked at the product name and its shortlisted photos together and picked one, with a score from 0 to 100. Low scores went to a review list instead of the shop.
Categories for every product
A text model sorted all 5,977 products into department, category and subcategory, ready for the online shop's menus.
Files ready to upload
Renamed photo files, contact sheets for the review list and upload packages, plus flags for products sharing one photo and images too small for the Meta catalogue.
How it runs day to day
A 200-product pilot came first, so the confidence bands could be checked by eye before the full run. The full catalogue then ran overnight. The shop's team worked only through the review list of medium-confidence matches instead of the whole catalogue.
Results
- products matched with high confidenceOutput mapping file, 5,977 rows
- 5,261
- matches sent to a person for reviewOutput mapping file, medium confidence band
- 607
- products sorted into department, category and subcategoryCategorisation output file
- 5,977
What's next
Re-running new products as they are added, so the online catalogue never falls behind the shop shelves.
Related
Industry
Let's put AI to work in your business.
One team designs, builds and runs all nine products. You talk to the people who wrote the code. Or WhatsApp us on +971 56 121 1011.