Label less. Learn more.

ALaaS looks at your whole image dataset and picks the most informative images to label - so you train better models with a fraction of the budget. Pay-per-use, API-first, built for agents.

1 · Upload

alaas upload ./images --pool <id> - your images are processed right away; the originals are never stored.

2 · Select

alaas select --pool <id> -k 50 - our selection returns the most useful images to label next.

3 · Repeat

Each round excludes what you already labelled - progressively covers your dataset.

Does it actually work? We measured it.

We ran controlled studies on real photos, satellite imagery and medical microscopy: pick images with ALaaS or at random, train the same classifier, compare accuracy at every labelling budget. 5 random splits per dataset.

25%50%75%100%04080120160labelled imagesALaaS selectionrandom sampling

Classifier accuracy vs labelling budget, natural photos (6 classes). Mean over 5 splits.

2.5x fewer labels

ALaaS reached 98% accuracy with 60 labels. Random sampling needed about 150 for the same result.

+11 points where it hurts most

At tiny budgets (16-40 labels) ALaaS beat random by 5 to 11 accuracy points - exactly the cold-start phase where every label is expensive.

Matches state of the art, without a model

Across three domains, ALaaS tied BADGE (the strongest published method we tested) on average - while needing no trained model and no existing labels to start.

Honest fine print: controlled study protocol, 5 random splits per dataset; gains are largest at small budgets and on data the service handles well. On far-out domains (e.g. blood microscopy) the advantage shrinks - we tell you that instead of hiding it.

Read the full studies

Why teams pick ALaaS

Works from zero labels

No model in the loop, no warm-up rounds. Point it at a folder of unlabelled images and get a useful first batch immediately.

Your images stay yours

Uploads are processed right away and discarded - we never store your images. Nothing to leak, nothing to delete later.

Pay per use

Credits, no subscription. Uploads and selection are metered per image; idle pools cost nothing.

Built for pipelines and agents

Plain JSON over HTTPS, a one-command CLI, and an OpenAPI spec your agents can read. Selection rounds are repeatable and never return the same image twice.

Start with 10,000 free credits

Enough to upload and select your first dataset. No card required.