Mariia Izofatova

Trained a BERT model from scratch for banner ranking via session-based next-click prediction (offline: Hit Rate@3, NDCG@3; converted to ONNX for production). Set A/B-testing criteria (CTR, conversion). Built offer-relevance models feeding a Gross Income equation; sped up scoring via PySpark (hours → 30 min). Wrote a greedy algorithm for GI-based client offers. Validated metrics for a RAG-based Text-to-SQL assistant's eval service (LLM-as-judge, Recall@15/Precision@15).

Fine-tuned internal BERT models to classify fraudulent search result pages (~0.98 quality), distilled into a DSSM, deployed to production. Built validation metrics and quality dashboards for a fraudulent phone-number detector. Worked with model quantization and distributed training (DP/DDP/FSDP).

Fine-tuned Gliner (NER) for client data anonymization, using an LLM for data augmentation. Designed a chatbot delivering personalized daily sales results to retail staff. Built and shipped a recommendation system for CVM subscribers (feature space + boosting model). Designed a CatBoost model ranking stores by predicted shortage, confirming fraud and delivering €600K in incremental income. Contributed to an internal ML platform. Led a team of 7 interns for 4 months.