Aleksandr Rezanov

AI video generation startup targeting on revolutionizing video generation process for marketing and consumer markets ● utilized post training of diffusion models for customized video generation achieving state-of-the-art motion quality according to the human evaluation; ● owned and developed fast custom pipeline for autonomous data preparation; ● introduced and developed a novel MLOps approach to reduce prototype time-to-market from weeks to days.

AI video generation startup (Brask) with video localization platform (Rask) ● developed novel GANs architectures to achieve state-of-the-art quality of lipsync in terms of mean opinion score; ● merged active speaker detection and speaker diarization to release first-ever multispeaker lipsync; ● built best-in-class out-of-sync detector leading to the increase in synchronization quality by 15% according to the human evaluation; ● developed GPU-only image processing pipeline to reduce video inference processing time by a factor of 5; ● adopted diffusion models to create a synthetic dataset to increase lipsync model’s synchronization quality by 15%; ● led a team of 5 Research and Machine Learning Engineers to develop one of the best lipsync solutions on the market.

As a Research engineer, I was in charge of developing a state-of-the-art lip-sync approach.

The leading AI engine for NPC development. Most funded startup in AI for gaming ● built a custom queue and register implementations to parallelize data processing and enable request deduplication, leading to the reduce in processing time by 50%; ● developed a probabilistic model to move from reactive agents to proactive.

AI consulting partnering and impacting top tech companies: Huawei, NVIDIA, Amazon, and Microsoft ● built model for generation of gym training plan using generative transformer beating quality of the top tier US coaches in the terms of mean opinion score; ● developed a robust pipeline for sota face mask detection on ip camera under challenging environment in the wild with detection mAP of 99% and classification accuracy about 90% (while available solutions had wild guess accuracy); ● optimized detection pipeline postprocessing by finding and reimplementing bottleneck leading to the reduction of processing time from 60 to 0.5 seconds.

Self-driving laboratory at MIPT ● developed a novel solution for CNN-based localization using GAN for bird-eye view synthesis; ● built fast and robust novel 6d pose estimation method based on CenterNet achieving 20 fps on single gpu while having sota quality.

Master program organized by one of the leading computer vision companies - ABBYY. Master thesis - research of transformers for document image dewarping