🔸 Core AI engineer🔸 Develop an engine for efficient inference of complex AI pipelines on videos🔸 Develop an end-to-end pipeline for generating DeepAvatar videos

🔸 Developed an engine for efficient inference of complex AI pipelines on videos🔸 Developed end-to-end pipeline for generating face swap videos: decoding, detection, tracking, face generation, blending, encoding.

🔸 Developed one of the most accurate face recognition systems in the world (confirmed by independent tests NIST FRVT, IARPA FRPC)🔸 Was a key developer of most components of the video analytics system: object detection, face recognition, pedestrian recognition, object tracking, and much more🔸 Developed the core of the AI system, which processes more than 100k real-time video streams🔸 Played a crucial role in organizing the work of the R&D team (from data collection to deployment to production)

🔸 Improved the time-tested OCR system by replacing the old-school algorithm with CNNs. Did all the work from research to production.🔸 Invented a convolution algorithm on an RLE-image, which is several times faster than the GEMM-based implementation. This convolution is the most expensive operation in the tiny neural networks that are used for OCR.🔸 Did research on the application of CNN + LSTM combination for Arabic text recognition. For training, I used only synthetic data with augmentations. The resulting algorithm works faster and more accurately on real data than the previous production version.🔸 Came up with a differential classification scheme to speed up character recognition. See patent "Differential classification using multiple neural networks"