Mostly connected with the Computer Vision Problem. Quality enhancement: video upscaling & denoising. Working with GANs and Diffusion Neural Networks. Implement and enhance architectures form papers and open-source solutions. Finetunning existing models: Lora adapters & DreamBooth techniques

First of all, Teaching assistant in Data Science subject for CS-students. Also Research in deep learning fields, mostly in NLP, RL and Time Series. Projects that have been done: 1. Working with LLM's, fine-tunning for specific task e.g. classification, summarization and question-answering problems. Using both BERT's and GPT's structures. LoRA tunning. Mostly working with LLaMa and phi-2 models.2. Created a classifier to identify obscene/prohibited comments with roc auc > 93% on validation3. Created a salary predictor based on textual descriptions of job location and responsibilities, using CNNs and various improvements: improving the CNN architecture (dropout, batchnorm, parallel convolutional layers), changing pooling, and moving to LSTM.4. Different RL-algorithms implementation from scratch (For example, Deep Q-Learning and its modifications: Hard target update, Soft target update, Double DQN) and solving different gym areas, more -> on my github. 5. Realisation of the whole pipeline time series forecasting problems, starting from econometric methods (smoothing, SARIMA, stationarity and seasonality studies), including classical ML - feature extraction, boosting, and ending with Deep Learning: Parallel RNN, LSTM and Transformers.

Experience working with large data sets, mostly in teams of 3-7 people. Designed and developed various mathematical models in Python describing experimental data (in physics, for example, significantly improved the SRH+ model).Major accomplishments include:1. Creation of an algorithm for approximating experimental data in Python.2. Development of software for obtaining stationary solutions of the system of kinetic equations.3. Modernization of the widely used SRH+ model to describe luminescence decay curves. The accuracy of the model has been increased by more than x2.Technology stack used: Python (Numpy, Pandas, Scipy, Sklearn, Seaborn & others), MySQL