Problem
A mobile bank saw that 30% of its transfers go to phone numbers that are not in the customer's contacts. Calls to customers showed that most of these are payments to small sellers without their own card acquiring: market stalls, small shops, manicurists. Typing a number takes about 20 clicks. The team wanted to cut this to a camera shot, with recognition accurate enough that nobody has to retype.
Example from the article: at a market stall that takes no cards, the customer points the camera at a printed or handwritten number, and the app offers a transfer in 2–3 clicks.
Solution in brief
A two-stage vision pipeline: a CRAFT-style detector finds the number, a CRNN reads it. The recognizer is trained mostly on synthetic data, including handwriting assembled from collected digit templates. After launch, low usage forced a rework of the entry point and onboarding, and the feature reached a 25% monthly retention plateau.
Questions this case answers
- How to replace manual typing of long identifiers with camera capture and reliable text recognition when the identifier appears in printed or handwritten form
- How to choose a detection quality metric that matches downstream text recognition accuracy rather than box overlap
- How to replace manual typing of long identifiers in a mobile app with camera recognition
- How to find which manual user action is worth automating using transaction data
Tools and methods
- CRAFT-style detector — finds numbers of any shape and angle
- CRNN with CTC loss — reads the cropped number
- Synthetic data — printed fonts and handwritten digits cut from templates
- Toloka — crowd photos of real numbers
- Optuna — tunes postprocessing
- Triton Inference Server — serves both models
- A/B test, retention curves — check zoom and long-term use
Google or a one-time email code. No password.