AI receipt scanner: turning a photo of a receipt into itemised data

Traditional OCR recognises glyphs and leaves the meaning to you. A vision model reads the receipt the way a person does — it knows a receipt has a merchant at the top, line items in the middle, a subtotal and tax near the bottom, and that a number in the right-hand column on the same row as a product name is that product’s price.

That structural understanding is what survives damage. A fold through the middle of a line does not stop a model that can infer the item from context and the price from the column. A price smudged past legibility gets flagged as uncertain rather than silently guessed, which matters far more than raw accuracy — a scanner that is wrong confidently is worse than one that is right less often and says when it is unsure.

It also handles the abbreviations that make grocery receipts unreadable. KS ORG CHKN BRST comes back as Kirkland Organic Chicken Breast, because a model that has read the internet knows what a Costco SKU looks like. No lookup table gets you there.

TabDaddy sends the photo to Claude, which returns the receipt as a strictly-validated structure rather than prose: one entry per printed line with its name, quantity, unit price and extended total, plus the subtotal, tax, tip, fees and discounts, all in integer cents. Nothing is parsed out of free text and no price is ever pulled out of a sentence with a regular expression — the model is constrained to a schema, and anything that does not conform is rejected rather than patched.

Then the server checks the model’s work against itself. The sum of the line items, minus discounts, plus tax, tip and fees, has to reconcile with the printed grand total. Within a couple of cents it is rounding and the receipt passes. Outside that, the receipt is flagged and you are shown where it does not add up, with the missing line highlighted, so you can fix it before anyone is charged.

It never quietly adjusts a number to force a match. A scanner that balances the books by inventing a value is how someone ends up billed for an item that was never on the receipt.

Extraction quality is mostly a photography problem. Five things account for nearly all failures:

Reading the receipt is the boring half. TabDaddy exists because the useful half is what happens next: you tap each item onto whoever ordered it, per-person totals build up with tax apportioned by subtotal, and each person gets a message naming their items with a Venmo link carrying their exact share.

Your first scan is free and takes no account. The photo itself is transient by default — it is sent, read and dropped, and is only ever stored if you explicitly opt into posting it publicly.