← Back to projects
Visual Product Search
Take a photo of a product and it tells you which one it is, among tens of thousands, and which stores have it in stock.
3capabilities: search, classify and locate stock
2search modes: photo and text
The problem
Identifying one product among tens of thousands from a photo was impossible without knowing the code by heart.
What I did
I built the catalog's multimodal search on the Lab's shared vector layer: 512-dim CLIP embeddings over Qdrant, category prediction from a photo and a store map with stock.
The outcome
In production as the catalog's search: by photo or by text.
How it finds the product
- A photo or a textInput
- CLIP embedding512 dimensions
- Similarity searchOver the vector store
- Category predictionFrom the photo itself
- Stock per storeOn a map
- Product identifiedOutput
A photo goes in and a product code comes out, among tens of thousands. The visual vector does the heavy lifting; the predicted category and per-store stock turn the find into something actionable on the floor.
Stack
- Python
- CLIP
- Qdrant
- PyTorch
- Streamlit
- Google Maps API
- SQL Server
Blue Star Group internal project — the code is not public.
Next project: Shared Vector Infrastructure →