Providers who deliver computer vision.

Computer vision is a class of system that turns pixels into structured predictions, bounding boxes, segmentation masks, class labels, extracted text, embeddings, or pose estimates, usually via a trained model (CNN, ViT, or a multimodal foundation model) wrapped in a pipeline that handles ingestion, preprocessing, inference, post-processing, and human review. It is not one tool; a defect-detection line, a document OCR service, and a retail shelf-audit app share the label but have almost nothing in common operationally.

When evaluating computer vision providers, ask for per-class precision and recall on a holdout that resembles your own imagery, not a public benchmark. Ask how the pipeline behaves on the long tail (occlusion, motion blur, unusual lighting, out-of-distribution classes) and how false positives are surfaced for review. For any visual AI touching people, ask what demographic slice testing has been done and how the provider handles the drift that arrives the moment cameras, lenses, or environments change.

Failure modes are specific. Demos on curated frames that collapse on real-world footage. Vendors quoting mAP without disclosing the eval set or the confidence threshold. CV implementation projects sold as a model when the real work is data labeling, edge deployment, and MLOps. And a persistent one: OCR pitched as "solved" when handwriting, low-DPI scans, or non-Latin scripts still miss badly.

Trustgent rates each provider on the L0-L5 verification spectrum, from self-attested claims up to independently reproduced evaluations and third-party audits, with the exact evidence viewable on each profile. See /how-we-verify for the level definitions and what a claim at each level requires.