← All projects

Automation & data

Built in January 2026

The customer describes the product in their own words

The engine finds the right item, from a text or from a photo.

An animated diagram, demo data: a text and a photo sent to the API, turned into a multilingual vector and a CLIP vector for the store’s pgvector search.
The animated diagram, demo data: the text and photo requests.
The animated diagram, demo data: models loaded, vectors computed.
The animated diagram, demo data: the embedding the API returns.

The problem

Customers describe a product in their own words or send a photo, and the agent must find the right item in the catalogue.

What I built

A FastAPI service that produces multilingual text vectors (768d) and CLIP image vectors (512d), plus image descriptions, for the pgvector search in the store’s database.

Technical details

Input validation with Pydantic v2
CLIP output handled across transformers versions
Batch endpoint for bulk embedding

Stack

Python
FastAPI
sentence-transformers
CLIP
Florence-2
Docker

The result

The embedding servers answer 503 until their models are loaded, so the host sends them no traffic before they are ready, and 4 container variants fit the hosting plan.

A similar project in mind?

Tell me about your business and when you need it. You get a proposal by email.

I want this system

More projects

Automation & data · Darija intent classifierA false confirmation ships a parcel nobody wantsView the project