denismagnitov

Won 1st Place in the ML/Web track by developing a high-precision credit scoring model for Alfa-Bank. Engineered a LightGBM-based model that outperformed CatBoost/XGBoost benchmarks, achieving a 24% reduction in WMAE and high stability (CV 3.15%). Implemented a robust training pipeline using K-fold cross-validation and MAE (L1) optimization to handle data outliers effectively. Developed 'AlfaScore,' a full-stack web service featuring an interactive UI dashboard and SHAP-value integration to explain model predictions to sales teams.

Designed and built a fully automated Telegram bot for selling and managing VPN access keys. Implemented complex business logic integrating with the 3x-ui API for user provisioning and traffic monitoring. Containerized the entire application using Docker and deployed it on a self-managed Linux VPS, ensuring 24/7 uptime. Designed a normalized SQL database schema to handle transaction logs, user subscriptions, and automated migrations.

Developed deep learning models using PyTorch/TensorFlow for multi-class vehicle classification. Built an end-to-end pipeline for dataset preparation, augmentation, and model validation using Accuracy/F1 metrics. Applied Transfer Learning techniques to optimize pre-trained architectures for specific vehicle recognition tasks.