Cogniscient
Applied ML research · Australia

Accuracy where the labels run out.

Cogniscient develops fine-tuning methods for vision models trained on a few hundred labelled images — the regime most real-world defect classification actually lives in.

60 70 80 90 Accuracy % +11 pts Cogniscient method Standard fine-tuning 50 100 200 300 500 1k 2k Labelled images per class
Shape of the result, drawn to illustrate the regime — not measured data. Per-dataset figures, baselines and the full evaluation protocol are available on request.

The regime

Every new customer, product line or inspection standard is a fresh model trained on whatever was labelled by hand. Usually a few hundred images.

The result

On public industrial defect benchmarks, 2–13 percentage points of accuracy over standard fine-tuning of the same backbone. Largest gains at 300–500 labelled images per class.

The offer

Free benchmark pilots under NDA for vendors already running defect classification in production. Your data, your splits, side-by-side against your current pipeline.