I've built production ML systems for Avito, including LLM-based support assistants that reduced median first-response time by 25%.
I develop retrieval, matching, ranking, and recommender pipelines, from hybrid semantic and BM25 search to RAG-based ranking. My work improved search relevance by 5% and 10%, while contrastive-learning embedding optimization reduced latency by 25%.
At Avito, I fine-tuned and pre-trained domain-specific LLMs with LoRA and QLoRA, reducing GPU-hours by 35% and improving target-task performance by 40%. I also established continuous training, deployment, monitoring, and A/B testing workflows with Airflow and Kubernetes.
Previously, I built Spark-based product matching and ranking pipelines at OZON.ru and improved Yandex Alice activation accuracy at Yandex, reducing False Rejection Rate by 2%.
