Playground
Run the real API on a photo and see what comes back. Our samples are hand-counted, so you can check the answer against ground truth — or upload your own and see how it does.
Try it
A real call against the live API — same endpoint, same models, same latency you would get. Pick one of ours, or upload your own.
Getting a good result from your own photo
- One row of stacks, side by side. Anything else is rejected with
422 invalid_framingrather than counted — the model is calibrated for that arrangement and returns a confident wrong number on others. - Centre the row and fill the frame. Every photo the model was trained on is composed that way. Shot from across a table, individual chips get too thin to separate and the count comes back low.
- Shoot roughly level with the table, keep the top chip's printed face visible and in focus, and avoid direct glare on it — glare is the biggest cause of a stack falling back to colour matching.
- List every denomination that could be on the table, and no more. Extra values give the reader more chances to pick wrong.
Your photo is sent exactly as you chose it — we do not crop, pad or resize it first, because re-framing changes how large the chips appear to the model and moves the count. It is not stored.