iOS + Android
launched in both stores
0
receipt fields typed by hand
Context
Romanian companies with vehicles have to keep a logbook — foaie de parcurs. It records trips, fuel, mileage, purpose. It is a legal obligation, it is tedious, and there is a small industry of software for it.
Every incumbent in that industry solves the same half of the problem. They make the calculation faster: totals, tax treatment, exports for the accountant. And they all assume a human sits down and types in the fuel receipts. The calculation was never the painful part.
Constraint
Removing the typing means reading a photograph of a crumpled thermal receipt taken at a fuel pump in bad light, and being right often enough that checking the output is faster than typing it. If a user has to verify every field anyway, you have built a slower version of a keyboard.
It also had to be mobile-first in the real sense. The moment a receipt exists is at the pump, not at a desk that evening, and a receipt that is not captured in that moment is usually lost.
Decision
We rejected competing on calculation. It is the feature every competitor demos, which means it is where you win nothing. Attacking data entry meant the product's core risk became OCR accuracy rather than business logic — a much less comfortable place to be, and the only differentiator worth having.
On the pipeline itself, we rejected passing the raw photo to a language model and asking for structured output. It works and it costs too much per receipt for a product priced for small Romanian fleets. OCR extracts the text, the model interprets and corrects it. Cheaper per scan, and when a receipt comes out wrong you can see whether the extraction or the interpretation failed.
On distribution, we rejected shipping a mobile web app first. A web app avoids two store review processes. It also degrades exactly where the product lives: camera access, offline capture, and a home-screen icon someone taps with one hand at a pump. We built native apps for iOS and Android and absorbed the review cycles. Mobile apps by Andreea Vișan.
What shipped
A Romanian SaaS for vehicle logbooks with iOS and Android apps live in the stores, built on Laravel and React Native with an OCR plus LLM scanning pipeline on AWS. The user photographs a receipt and the logbook fills itself in.
Numbers
iOS and Android apps launched. Manual entry of receipt fields removed from the core flow. Positioning built on eliminating data entry rather than optimising the calculation everyone else optimises.
What I'd do differently
I would have collected a labelled set of a few hundred real Romanian fuel receipts before building anything. We tuned against receipts we happened to have, which skewed towards a handful of chains, and the long tail of small independent stations cost us accuracy work later.
I would also have designed the correction interface earlier. When scanning is right 95% of the time, the entire user experience is decided by what the other 5% feels like, and we treated that as a detail for too long.