• Extended existing pipeline from single class to multiclass. This helped to decrease cost of inference by 20%. • Created grid search pipeline for existing models, which helped to increase accuracy of the models for various classes by 8%. • Suggested semi-suprevised learning to highly decrease labeling time and increase labeling speed. This feature allowed to utilize big amounts of unlabeled data and increase the accuracy of the whole pipeline by 6% without labeling efforts.

Historical texts OCR. Worked tightly with historical department and translated their business needs to code. During the project I gave the instructions for labeling, training the OCR model, collected metrics, feedback and created image-to-text system, which allowed increase the speed historians are working with. MVP for video interview emotion recognition and text-to-speech analysis. I created the prototype of a streamlit service for emotion recognition and TTS analysis for HR systems. My work helped to decrease the time of processing candidates’ interviews by a hiring recruiter thrice. Multimodal emotion recognition using attention in the wild. During the project the model for challenging in-the-wild dataset was created, final accuracy of the model is 62%, which is only 2% lower than state-of-the-art model. The model was optimized, using Apache TVM. Optimized model reduced costs by half and increased model inference speed by 30%

Steel anomalies detection. I Developed the model and made a service for it. Product, based on this service helped to reduce production costs twice decrease the level of defects in steel production by half. Streamlit service for Reader-Translator-Generator I made Streamlit service, which helped to further cooperate on the project and make the end product using provided codebase.