Our machine learning dev team developed a complex of neural networks that solved the search and recognition problem relating to the identification of car number plates in a video stream. It was necessary to achieve consistent high-quality results with more than 90% accuracy and to work with a 4k video stream at speeds of 60 FPS upwards. In addition, the software complex was required to operate on an embedded computer NVIDIA Jetson TX2. This required maximum acceleration and optimization of neural network calculations. In this project, our machine learning dev team not only developed and optimized the software complex but was also responsible for the collection and markup of training data (more than 20 thousand images), which was provided to the client.