Case study · Textile manufacturer · name withheld
Point a camera at a saree. Find it.
A leading textile manufacturer needed two answers fast: which catalog design is in this photo, and what do these marketplace settlement files mean. We built a visual search engine and an AI reconciliation layer that answer both.
384
Embedding dimensions
8
Vectors per image
6
Filter tiers
2
Systems, one platform
Station 01 · The problem
Which design is this?
A large catalog of designs, and a photo on a phone. Matching one to the other by eye takes minutes and gets it wrong. Meanwhile, every marketplace sends settlement spreadsheets in a different shape, and someone has to decode each one.
Station 02 · The search
The camera finds the catalog.
A photo is cleaned of its background, cut into eight crops, and embedded by a vision model into 384 dimensions per crop. A vector database matches against the catalog through six tiers of attribute filters and returns the closest six designs.
Station 03 · The reconciliation
Settlement files, decoded.
Every marketplace names its columns differently. The system retrieves the closest known format, hands that context to a language model, and gets back exactly which columns matter for reconciliation. New formats stop being a research project.
Station 04 · The evolution
From experiment to instrument.
The first version proved the idea with a heavier model running locally. The shipped system re-architected it: a lighter ONNX model, eight vectors per image, cloud vector search, authentication and a real deployment. Built to be owned, not demoed.
System map · verified
The architecture, on paper.
Users
React 19 app
camera + uploads
Core
FastAPI core
DINOv3 ONNX · rembg · 8 crops
Data
Qdrant
8 vectors / image · 6 tiers
Supabase Postgres
users · mappings
Outside
Groq LLM
column detection
ImageKit CDN
catalog images
Arrived. Now let's chart yours.
Every client gets a new flight path. Bring us the destination and we will engineer the route.