Vision systems that answer
operational questions.

Counts objects crossing a line by class and direction, measures how long each lingers in a zone, and estimates queue wait using Little's Law — which needs only counts, so it survives crowds where tracking degrades. Hysteresis eliminates the double-counting that makes naive counters unusable: on 400 frames of a track jittering on the line, a naive counter produced 196 false counts and this one produced zero.

Running a reasoning model on every frame of one camera costs more per day than the camera earns in a year. So a cheap detector watches everything, arithmetic rules nominate moments worth a second look, and segmentation plus a vision-language model run only on those — with the model's output forced into a fixed schema so a hallucinated field raises an error rather than an alert. Configuration alone moves the cost by 7×.

Extraction is the easy half. The half that makes it a product is knowing when the extraction is wrong — because a wrong total entering a ledger costs more than the automation saves. An invoice is an arithmetic object: line items sum to the subtotal, subtotal plus VAT equals the total. Those identities are free ground truth on every document, so the system can prove its own answer before posting it unsupervised.

A classifier needs examples of every defect, and a factory has thousands of good photos and perhaps a dozen bad ones — of the defect types that already happened. This inverts the problem: model normality in fine detail, flag anything that departs from it. No defect is ever labelled, and novel fault types are caught as a consequence rather than an extension. Results match the published benchmark.

Image collections are searchable only to the extent that somebody described them — and that description is fixed at the vocabulary chosen when it was written. This embeds images and text into one shared space, so a typed English sentence finds pictures nobody captioned, and the searchable vocabulary is open-ended. A controlled comparison here showed query phrasing alone is worth 14 accuracy points.