At Kontur, I built an LLM-based Text-to-SQL application for Russian business questions across 36 SQL Server databases. I took it from prototype to a two-week pilot with 10 employees, using retrieval and SQL validation without execution.
I treated evaluation as part of the product: the pilot answered 48 of 60 questions with SQL, and 31 needed no edits. At Magnit, I updated a demand forecasting model and reduced WAPE by 5.3% versus the previous approach.
At Chatium, I shipped purchase-propensity scoring for call-center prioritization, increasing call-center efficiency 3x. Earlier, at Yandex, I built a personalized push-campaign system and an automated landing-page generator from search logs; I also hold an MS in AI from ITMO.
