Projects

Real-time Detection & classification road deffects

Developed a CV system for detecting and classifying potholes in road surfaces in real time using a moving video stream. Designed an inference pipeline optimized for working with dynamic scenes and varying road surface quality.

Python Computer Vision YOLO (v8/v11) Object Detection PyTorch

Developed a unified architecture for speech processing (TTS & STT) based on transformers and variational autoencoders (VITS-like)

Designed and implemented an end-to-end speech synthesis and recognition model based on the VITS architecture from scratch. Implemented the full cycle, from data collection and preprocessing to writing custom neural network layers and setting up training pipelines.

PyTorch Deep Learning NLP Audio Processing Neural Networks

Visual monitoring of construction equipment

Trained and deployed an object detection model for tracking construction equipment on site. Resolved issues with the model's performance in challenging weather conditions and when objects were obscured.

Python OpenCV Object Detection PyTorch Neural Networks YOLO

School Food Quality Control System (implemented in the Moscow region)

Created an end-to-end CV solution for analyzing tray photos and recognizing dish ingredients. Developed a Telegram bot for collecting feedback and integrating the model's predictions with the user interface. The project is successfully used in the real sector.

Python OpenCV Object Detection PyTorch Neural Networks YOLO Telegram Bot API

Predicting the market price of rebar for the company: [Samolet](https://samolet.ru)

Developed an ML model for forecasting the market price of rebar based on time series analysis and historical procurement data. Implementing these forecasts allowed for optimization of the procurement process and increased margins by selecting the optimal transaction window.

Python Gradient Boosting (XGBoost/CatBoost) Time Series Analysis Pandas

Classification of physical activity patterns (IoT Data)

Developed a data clustering pipeline for wearable sensors to determine physical activity type. Conducted feature engineering, configured KMeans clustering algorithms, and validated the results using Silhouette Score and the elbow method.

Python Scikit-learn Clustering NumPy Seaborn