Street Food Shuffle is my card game about Indian street food. Scoring a round means counting combos across a messy table: a Vada with a Pav, three Manchurians in a stack, a Chutney that doubles whatever it sits on. It is the slowest part of the game, so I built a scorer for it. You take a photo of your plate and the app adds up the points.
It is live at https://streetfoodshuffle.online/score.
How it works
- Everything runs on the phone. The card reader runs in the browser. No photo is uploaded, and no AI service is called during a scan.
- A detector finds the cards. A D-FINE model draws a box around every card, including the thin strips that show in a fanned stack.
- Two small checks read each card again. One tells the two look-alike puri cards apart, and one reads every card type from a close crop.
- The rules engine scores the plate. Which cards touch, in which order, decides the combos.
- The player always confirms. The result is shown card by card. When the reader is unsure, it asks a short question, like "Is this a Chutney?", and only the player's tap makes a score.
How I tested it
I trained on 228 photographed plates and set every threshold on those alone. Then I froze the whole pipeline and scored it once on 200 fresh photos from two new game sessions. The labels for those photos were written from the photos alone, without seeing what the reader said, and I checked every one by hand.
On that fresh set the reader got the score right on 86% of plates and read 83% of plates completely right, with nothing to correct. About one plate in six needs a correction, and most of those come with a question that points at the problem.
What is next
The misses cluster in two places: long fans of puri cards, where only the bottom strip of each card shows, and cards almost fully hidden under a stack. Both are rare in the training photos, so the next round of training adds exactly those.