We were the first in the country to create a selfie-based onboarding system for a bank using computer vision (machine learning and image processing AI systems).
The financial institution wanted to be the first in the country to enter the market with a non-real-time customer identification (self-service) system. As the regulator of the banking system (the Hungarian National Bank) had only issued a draft of the regulation that would technically enable this solution at the time the project started, working closely with the regulator itself was critical to the development and interpretation of the proposal itself. Implementation required a new module in the mobile app that would enable this, while fitting organically into the existing logical structure of the app.
We started the implementation on three fronts. Our legal team consulted with the regulator on the details of the draft and channeled to the regulator the technical limitations we felt were important to the proposal. Our business analysts and UX team worked out the most natural way to integrate the new module into the existing system. Meanwhile, our data analytics team used a computer vision-based AI to build an algorithm that exports the user's face from the frames in the selfie video, identifies whether it matches the face on the uploaded ID card, and tests the liveness of the person in the video.
Thousands of accounts have been opened through our selfie-based account opening solution since its launch, with the number of daily account openings on this channel now matching the other most successful channels. Our AI algorithm, which performs face matching and liveness checks, has a 100% success rate, with no false positives reported
This project perfectly illustrates our approach to bring an MVP product to market as soon as possible, deploy metrics and measure back how users use the product and achieve incremental results with quick optimization at least as much as the product launch itself.
In the case of selfie account opening, we have found that a large percentage of customers take a photo of the wrong side of the address card, so the account opening fails. For the first time, we improved this metric by inserting a picture of the correct side of the address card before the photo was taken to illustrate which side was needed. While the success rate improved, it was still not high enough, so we developed a character recognition (OCR) algorithm to give real-time feedback to the customer who had taken a photo of the wrong page, reducing the chance of error to 0%. This increased the number of successful account openings by 50%, effectively doubling the ROI index of the project.