Retina: vision models without the MLOps yak-shave

Most computer-vision projects die in the plumbing: object storage, training infrastructure, experiment tracking, serving. Retina packages that plumbing so a vision project starts at the interesting part — the data and the model.

Groups, uploads, training

Images are organised into groups and uploaded in bulk to object storage. Training runs execute on background workers with live progress in the UI, so "is it still training?" has a visible answer rather than an SSH session.

Experiments you can compare

Every run is tracked in MLflow: parameters, metrics, artefacts. When you have trained five variants, choosing a winner is a table sort, and reproducing it later is possible because the details were recorded when they happened.

Serving is the product

A model that is not behind an API is a science project. Retina serves predictions through an inference API documented with OpenAPI, so client integration is a generated SDK away. Self-host the whole platform and your images never leave your infrastructure.