Case study · Jewelry · #ProjectBanana
A catalog that did not exist
In short
A jewelry business had no product data to clean: a register snapshot that no longer matched reality, scattered sheets, and product photos with nothing attached. alfa built the catalog from the photographs, with vision models that classify stone, metal, texture and shape, into one master of models and variants linked to the ERP in both directions. That master feeds the online store and the point of sale: 41,178 SKUs, 867 models and 1,743 product pages built.
Counts as of 28.10.2026.
The problem
The information did not exist. Not messy: absent. A register snapshot that no longer matched reality, scattered sheets, and product photographs with no data attached. By any current standard the catalog was analogue, and there was nothing to clean.
What alfa built
- Vision models on product photos (beta)
- Image-to-variant mapping
- Two-way ERP sync
- Auto export to POS
- Hebrew localization
The catalog, read from the photographs
Vision models read every product photo, and vector embeddings separate the features: stone, metal, texture and shape, each classified and mapped to the record.
A two-level master
Models on top, variants underneath, with each photo mapped to its variant. 867 models and 41,178 SKUs sit in one master.
Linked to the ERP, the store and the register
The master is linked to the ERP by direct API in both directions. The online store and the point-of-sale export are fed from it automatically. Hebrew localization is part of the build.
The result
One master that feeds the online store and the point of sale automatically: 41,178 SKUs, 867 models and 1,743 product pages built.
Vision checks on harvested images are described in every image on the right variant. Every file passes the validation gate before it reaches the store.