AI Platform Supporting
Sustainable Recycling

We focused on releasing MVP on the market in a short time to make sure the app was relevant to the market, as well as to build our further development roadmap based on users’ feedback. In particular, we have used pictures from users’ reports to prepare a custom dataset of over 2000 objects, which we utilized to train our model further. In terms of tight deadlines, the combo of Flutter & Firebase was optimal so that we could release the app for both iOS and Android simultaneously, without the need to write code separately for each platform or to build a custom backend.

After considering various ML models for the Recycling Assistant, we opted for YOLO since it fulfilled our key requirements, which were:

• YOLO is one of the fastest object detection algorithms currently available on the market.
• It achieves high accuracy in object detection, even for small objects.
• The model can detect a wide range of object categories as well as multiple objects in a single image.

Our decision to opt for YOLO5 instead of YOLO7 was based on several factors. It processes information at a faster rate, occupies less storage space on devices, and is currently more stable when used with Flutter. Although YOLO7 is a more powerful model that provides slightly higher accuracy, its longer processing time and, more importantly, larger size, were critical considerations. As the AI model performs calculations directly on users’ smartphones, we aimed to minimize the storage space required for the Recycling Assistant.

Close collaboration between our Quality Assurance and Data Science teams during manual testing resulted in the detailed report identifying objects that the model struggled with, such as mistaking cherry tomatoes for bottle caps or brightly branded paper glasses for soda cans. This helped the DS team to effectively plan and accomplish further model training activities.

To ensure a smooth user experience from the outset, the QA expert highlighted important external factors to consider before scanning, such as camera clearance and specific lighting and background requirements. These details are presented upon launching the app, helping to create an intuitive user flow.

To streamline the object CRUD process (create, read, update, and delete), we developed a multi-region admin panel that provides information on all categories and rules for each country we operate in (currently, the app covers Germany, Czech Republic, Poland, and the United Kingdom). Furthermore, the admin can access statistics on the most frequently scanned objects, as well as detailed bug reports with error information.