Splitty
A pile of receipts in, one clean PDF each out — and the files never leave the accountant's browser.
Built for the week before a VAT deadline.
Every VAT cycle, a French accounting practice receives the same thing from dozens of clients: one file with everything in it. A phone photo of five receipts spread across a table, one of them upside down. A four-page PDF scanned in a single pass.
Splitty takes that file, finds each receipt inside it, straightens and crops it, reads the text, and gives back one clean A4 PDF per receipt — named and numbered, ready to drop into Pennylane, Sage or Dext.
The work nobody bills for.
Splitting that file by hand is pure loss. It is not accounting, it is not advice, and it cannot appear on an invoice — but it has to happen before any of the real work can start, and it happens again every single month.
There was a second constraint that shaped everything else. These are accounting documents: bank details, client names, amounts. Any tool that quietly uploads them to a server to do the clever part is asking a regulated professional to take a risk on their clients' behalf.
Do the hard part on the device, and never export behind the user's back.
Detection, cropping and text recognition all run in the browser. Nothing is uploaded to be processed. That is a real constraint rather than a marketing line — it ruled out sending pages to a hosted model, and pushed the detection toward computer vision that could run locally and predictably.
The interface follows the order the work actually happens in: you drop the file, Splitty shows you what it found, you uncheck anything that is not a receipt and rotate anything it got wrong, and only then do you export. The validation step sits deliberately between detection and output — nothing leaves the tool without being looked at first.
Live, and free to try without an account.
Splitty is in production at splitty.fr. The free tier runs without a card so a bookkeeper can test it on a real client file in the time it takes to make coffee, and the paid tier is a flat per-seat subscription rather than per-document pricing.
The decision I would defend hardest is the one that is invisible when it works: keeping the whole pipeline on the device. It made the engineering harder and it is the only reason the tool can be handed to someone who is professionally responsible for the confidentiality of what is in those files.
