Sunoa.
A camera point-of-sale for a Kona patch shop: photograph the customer’s tray, the cart fills itself from hundreds of SKUs, the cashier confirms, and the terminal takes the tap.
Hundreds of SKUs, rung up by hand.
At the counter, a finished bag or keychain carries dozens of small individually priced embroidered patches plus custom lettering. Ringing one up meant hunting a catalog of hundreds of items and hand-counting every patch and letter, on every sale, with the customer watching.
Photograph, verify, charge.
The scan screen
One photo of the tray and the cart populates itself, with quantities editable and low-confidence matches flagged for the cashier's eye before anything is charged.
The vision pipeline
Frontier segmentation finds every patch and letter in the photo, embedding search matches each one against the catalog's own reference images, and a dedicated letter-reader counts custom glyphs by color and style.
Square, end to end
Matches become a priced, taxed order in the shop's existing Square catalog and land on the paired terminal for the tap. Built and proven against the sandbox; going live is a credential swap, not a rebuild.
The answer key
Accuracy is not an adjective here. A hand-verified ground truth of 279 labeled instances across 26 photos scores every pipeline change on precision, recall, and F1.
Sixteen logged experiments
Detection models, embedding models, judges, and thresholds were swapped and scored in a written experiment log, each kept or discarded strictly by the numbers.
The correction flywheel
A review workbench turns every cashier correction into new verified ground truth, so the system's answer key, and its accuracy, grow with use.
How we built it.
The pipeline composes frontier vision models for detection, image-embedding retrieval against the shop's own catalog photography, and a model judge reserved for the specific confusions the experiment log exposed. The discipline is the point: nothing ships on vibes. Every candidate improvement runs against the answer key first, and the design keeps a human in the loop, the cashier confirms every cart before a dollar moves.
Results.
88% of patch instances identified correctly on the current 213-instance evaluation, measured, logged, and reproducible, with the correction workbench feeding the next point of accuracy. The checkout that used to mean hunting hundreds of SKUs now starts with one photograph.