Data Scientist

Led development of an LLM evaluation library
Reduced development and testing time by 60%, accelerating product releases
• Deployed LLM-as-judge solutions
Saved up to 50% of human labor costs
• Integrated RAG systems and chatbots into model validation workflows
Reduced validation time by 35%
• Researched and implemented advanced LLM solutions
Implemented agents, multi-agent systems, and PEFT fine-tuning pipelines using LoRA
• Collaborated with cross-functional teams to implement AI technologies in production
workflows
Improved adoption of LLM-based tools across validation and development teams

Designed and implemented a testing library for pricing models
Accelerated product releases by 50%
• Identified and resolved critical gaps in risk metrics for exotic products
Prevented significant financial losses
• Developed predictive and regression models for traders and analysts
Increased trading accuracy and profitability
• Automated hedging error calculation for trading book products
Improved risk-adjusted decision-making for traders and analysts

Developed ML-based lead ranking models
Prioritized high-conversion leads across multiple sales funnels and customer stages
• Built calibrated scoring pipelines for different lead lifecycle stages
Unified heterogeneous model outputs into a single interpretable business score
• Improved sales funnel efficiency through lead prioritization models
Reduced lead-to-conversion time by 20%+ and increased conversion rate by 0.4 p.p.
• Built LLM-based lead enrichment pipelines for real-time feature extraction
Improved PR-AUC by 10%+ at final lead evaluation stages
• Designed model-driven analytics baselines for conversion monitoring
Helped teams detect conversion drops, compare sales performance, and explain funnel dynamics

Financial sector risk assessment. Developed adaptive ML pipelines for crypto token prediction dynamically selecting CatBoost, RandomForest, and LSTM models improving ROC-AUC by 7%. Optimized classification and ranking models for token risk assessment increasing ROC-AUC and F1 scores by 7%. Built automated alerting and retraining system to detect model degradation reducing manual adjustments by 40%. Enhanced backtesting framework for crypto forecasting models improving predictive stability by 5%. Collaborated with product and marketing teams to align ML outputs with business goals contributing to 18% LTV growth and 15% higher customer satisfaction.

Faculty of Mechanics and Mathematics