- Lead a highly skilled data science team of 7 in a subsidiary of a bank with over $40 billion in revenue, overseeing the development and deployment of advanced ML/DL models, ensuring alignment with business goals, mentoring and teaching, hiring new team members and fostering a collaborative environment to drive innovation and excellence in data-driven solutions.- Developed, fine-tuned and adapted Generative AI (GenAI) and Computer Vision (CV) models for our marketplace, including largest LLM for Russian language - ruGPT 13B (DeepSpeed sparse attention and parallelism, position of normalization layers, rotary positional embeddings, xformers), ruDALL-E Kandinsky, Summarizer, Upscaler, Object Detection and Colorization using Christofari supercomputer (V100, A100 gpu). Brought platform from 0 to 30k+ concurrent users.- Created various ML and DL models (PyTorch) for customers in finance, oil & gas, retail, and agriculture, generating over $25 million in profit. Such as predicting component failures from sensor readings (LGBM&CatBoost) or coal quality detection (EfficientNet B6)- Developed and customized 50+ Docker images with Kubernetes, foundational to our Cloud Inference and Notebook services, handling everything from selecting base images to implementing efficient Docker caching and designing complex layer architectures.- Worked on development and hosting competitions on Kaggle like platform DSWorks. Hosted such competitions as Sber RecSys Benchmark and AI Journey Contest with more than 2000 participants.

- Developed Credit Risk Scoring ML Models (CatBoost, LightGBM) using Qiwi's proprietary data from multiple sources (Oracle, Vertica). Utilized Apache Spark and Hadoop for data processing and access, including data cleaning, feature engineering, and validation. Created blended models incorporating various algorithms and styles, resulting in models sold to major banks in Russia, yielding over $10 million in profit.- Performed Analytics which improved client retention rates and reduced churn by 15%.- Created a Fraud Detection Model by generalizing patterns across different groups to identify fraudulent activities.

- Performed Sales Forecasting using a blend of LGBM and XGB models. These models were a critical component of a complex pipeline for predicting product purchase quantities by businesses, resulting in approximately a 10% annual savings on procurement costs.- Identified underperforming products using Matplotlib and Seaborn.- Collaborated with data engineers to optimize data storage and update solutions in PostgreSQL.