Dmitry Protasov

- Optimised item2item recommender: moved to HNSW vector search, split GPU/CPU stages; cutting compute cost by 30% and time inference 3x faster - Developed a Metric Learning-based vision model combined with CatBoost to compare cars in photos and detect fake listings, achieving 55\% recall on a specialized dataset - Built item2item clothing recommender service, suitable for style (FashionCLIP, Faiss HNSW retrieval, catboost reranker). Improve the quality of model with Visual Language Models - Implemented a low-latency Transformer-based recommender that serves several thousand RPS; Ran a bunch of experiments that boosted views / contact requests by +4% - Improved vin2param service: add Airflow‑driven PySpark ETL, MLflow registry, real‑time Postgres updates, and Grafana monitoring - Implemented RAG search engine for complex automotive queries with CLIP embeddings in Qdrant; improved diversiry and long-tail recall@10 +7 pp, lifted search CTR +3%

Developing a telegram bot @Audio2MIDIBot that translates any music into a musical score for playing on the piano

- Developed Graph Neural Networks (GNNs) with Transformer architecture for history handling + SAC/TQC/PPO RL policies for molecular optimization → score +15\% vs DFT baseline - Contributed to the open-source Schnetpack library and refactored the code to enhance performance

Improved Yandex Search Engine with NLP-transformer‑based spell‑checker, reduced uncorrected typos 37%. Added auto‑translation to English for tech terms; boosted click recall +6pp

Completed the full ML curriculum at YSDA, covering Computer Vision, Natural Language Processing, Reinforcement Learning, Deep Learning theory, Recommender Systems, Speech Processing, and Self-Driving Cars. Gained hands-on experience with efficient model deployment techniques such as PyTorch Distributed, quantization, OpenVINO, and TensorRT, and studied modern generative approaches including GANs, Normalizing Flows, Diffusion Models, and VQ-VAE.