The main task for us was to introduce a robust ML solution that will help assess content quality at scale and accelerate the process of onboarding new clients. We identified the main bottlenecks: the data gathering process, ease of introducing new evaluation rules, and speed of asset processing.

Our work focused on redesigning the architecture, allowing existing models and algorithms to be reused and run on a large scale. Part of the solution was introducing ML-focused microservices with better-defined responsibilities. By doing so, we were able to more easily auto-scale with an independent group of computing endpoints in Kubernetes.

We addressed the challenge of detecting undefined, custom image parts (such as logos, fonts, and particular objects). In addition, we created a framework for generating new training datasets (using data augmentation and synthetic data). The framework accelerates training and provides agile and robust solutions for client requirements. New, artificially created data solves the problem of detection.

We also suggested a messaging system for running most machine learning and computer vision tasks by multiple Celery workers in parallel. It allows computing each “rule” as a separate task and processing the entire asset analysis pipeline, significantly improving performance.