I build agentic and RAG systems and check that they actually work. For three years I have shipped LLMs to production in banking, ads and EdTech: GigaChat search and assistants at Sber, self-hosted Qwen pipelines and a profession recommender serving Moscow's employment centre. Most of my work is evaluation: LLM judges, red-teaming and statistics that hold up on small samples.
paper-lab |
book-translate |
ru-genealogy-search |
- Sea of Data · LLM / AI Engineer · 2025–now. EdTech startup, AI career guidance for schools, universities and employment centres. Evaluation of a production AI tutor: a self-hosted Qwen judge scored 1,245 tutor turns on 18 pedagogical dimensions against 262 lessons of award-winning teachers. A profession recommender and learning-path builder in production for 3,600+ users, Moscow Employment Centre and Moscow Polytech.
- Sber · LLM Engineer · 2024–2025. LLM-as-a-Judge evaluation for RAG that cut evaluation time by 35%; GigaChat query understanding for restaurant search; LoRA fine-tuning of GigaChat for marketing push notifications.
- Yandex · Data Science Intern, Ads · 2023–2024. Ad ranking research, A/B tests, linear models in production.
- AI tutor and its evaluation methods · Data Fest 2026 · video
- Automatic RAG evaluation systems review · Data Fest 2024 · video
- Tune in to RAGAS and customize RAGAS for yourself · Habr, Sber blog, 2024 · article
Python · PyTorch · LangGraph · vLLM · Qwen · GigaChat · Claude Code · FastAPI · Docker · DuckDB · MLflow · DVC
ACM ICPC semi-finalist, informatics olympiad coach.




