Project on Optical Recognition of Parking Space Occupancy on a Device with Limited Computing Capabilities: - Weekly calls with the client in English - Proposing approaches to problem-solving considering business requirements - Researching, preparing, and generating datasets - Selection of architecture, training, quantization, deployment of neural networks on a specialized neural processor - Developed a model that achieved 98% accuracy in real-world conditions (the requirement was 85%) Project on Compression of Surveillance Camera Video Streams: - Proposed a concept that allowed for a x5 reduction in stored volume for the h.265 codec, with only a 5% drop in quality - Training and optimization of neural networks Gesture Control Project for Smart TVs: - Optimization of model architectures for computationally weak processors - Accelerated inference by 27% with only a 3% drop in accuracy Main libraries/technologies used: pytorch, tensorflow, opencv, openvino, numpy, pycharm, jupyter-lab.

- Segmentation of document fields, OCR. Implemented a neural network model of printed text recognition, superior in quality to EfficientNet. - Face recognition, liveness detection (face anti-spoofing). - Creation of microservices for computer vision tasks (python, kafka). Technologies and libraries used: tensorflow, jupyter-lab, detectron2, opencv, dlib, numpy, scikit-learn, pycharm.

Project in the field of computer vision. Implemented a neural network model to remove shadows and highlights from photographs of faces - for training, I created a synthetic dataset based on restoring a 3d head model from one photo and software 3d rendering. Implemented a neural network model for correcting white balance and light levels in photographs - for training, I created a synthetic dataset based on my own algorithm for recovering from color targets. The results are superior to those of "Deep White-Balance Editing" (https://arxiv.org/abs/2004.01354) An algorithm was proposed for quickly determining the dominant color in an image, surpassing K-means in accuracy and speed, the algorithm was accepted for work. Implemented neural network segmentation of color targets in photographs, bringing them to a rectangular form based on affine transformations. Implemented neural network segmentation of facial parts. Implemented the functionality of saving/retrieving hidden textual information in the scales of the neural network. I was also involved in markup and preparation of datasets. Technologies used: python, opencv, tensorflow, jupyter, numpy, pandas, sklearn, docker.

Projects: 2020 February - 2020 October A software package for visual detection and face recognition of store visitors. Produced R&D and implemented the frontal part of the project in the conditions of hardware configuration limitations - software interaction with several local (usb) and network (rtsp) cameras simultaneously, neural network face detection in real time, selection of the best face photo from the sequence, sending it further along the pipeline. Created custom bootable installation images for booting from a flash drive and installing software on terminal equipment with minimal staff involvement. Technologies used: python, dlib, openvino, opencv, HTTPServer, threading, REST API, intel neural compute stick 2, google coral tpu. 2019 December – 2020 March. Distributed hierarchical multi-screen software and hardware complex for displaying advertising media materials (slides, audio, video) according to schedule, playlist, screens. Participated in the discussion of the system architecture, proposed a number of architectural and technological solutions. Implemented that complex. Technologies used: python, Django, raspberry pi. 2019 September - December. Creation of a software package for the Pepper robot (https://en.wikipedia.org/wiki/Pepper_(robot) ), which implements human-friendly communication with a person. Human-friendly means a combination of verbal (voice) communication, non-verbal (gestures), eye contact (tracking a person's movements), as well as proactivity. Participated in the discussion of the general concept, proposed a number of technological solutions. Developed architecture, implemented a software package, including: 1) The robot's story about himself and his abilities. 2) A story about the current weather in any city in the world. Tracking the weather in Moscow, proactive notification of the interlocutor about a change in the current forecast. 3) Gestures and movements by voice command (“High five!”, Dancing, “I want to take a selfie with you”, “Follow me”, etc.) 4) A story about the company's products in accordance with the concept of "sales funnel". Including simultaneous presentation of slides and videos on the tablet built into the robot. 5) Issuance of a virtual accumulative discount card to the client. Includes identification of the client by his phone number, or the introduction of a new client in CRM (surveying the full name with confirmation, sending an SMS with a code to verify the number, etc.) 6) Proactively informing others about new sales records or major purchases in chain stores. Technologies used: python, Choregraphe, naoqi, google cloud speech, google cloud streaming STT, REST API, pymorphy2, dialogflow, mqtt, mysql, sqlite. 2019 June - August. Study of the possibilities for Russification of the software and hardware complex "Pepper robot" (https://en.wikipedia.org/wiki/Pepper_(robot) ). Produced R&D, implemented software modules for cloud recognition (Speech-To-Text) and generation (Text-To-Speech) of Russian speech. Wrote demonstration programs in which the robot performs simple physical actions on a voice command in Russian (nodding the head, raising/lowering the arm, etc.); by voice request, search in Wikipedia and voice short information from the found article; perception of a telephone number by ear, obtaining information about the client from the corporate CRM system and its voicing) Technologies used: python, Choregraphe, google cloud speech, speech recognition. 2019 April - June. Automation of data collection from open sources on the Internet. ETL, scraping, data cleaning, data matching. Technologies used: python, pyscrapy, numpy, pandas, google cloud, multiprocessing, jupyter.

Programming in Matlab for the processing and analysis of electroencephalograms, programming of software and hardware complex for experiments, conducting experiments.

Business Informatics, Management in the field of electronic business and Internet projects

Instrument-making, Automatic Control Systems

OpenDataScience society, Machine Learning course

coursera.org, Convolutional Neural Networks

coursera.org, Neural Networks and Deep Learning

coursera.org, Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization

coursera.org, Structuring Machine Learning Projects

coursera.org, Введение в машинное обучение

Stanford University, CS224d: Deep Learning for Natural Language Processing

coursera.org, Machine Learning

coursera.org, Neural Networks for Machine Learning