The problem was to bring one of the world's largest image recognition ML models to the hands of as many people as possible to enable them to identify plants in real time. Build a global map of every plant on the planet.​ The solution came in the form of designing and building a custom native application and leverage a backend infrastructure to support it. The result was great! More than 30 million installs and an average rating of 4.4 globally. Nearly 15 million user plant entries and more than 100 million plant identifications. We started by building a custom camera solution for taking photos and preparing them for image recognition. We added some guides for the users and this worked well initially. But as the image recognition model was getting more complex, it was becoming more difficult to match the user images to the ML model. It was incredibly challenging and we tried multiple approaches, but the one that actually made a big difference was integrating a custom plant detection model, which runs on device.