Middle Data Scientist
JSC Bereke Bank
A commercial bank operating in Kazakhstan.
Jan 2026 — PresentBuilt and deployed fraud detection, fraud segment discovery, and pre-default scoring models, including a report-generation service that cut analyst report time to seconds.
• Built a CatBoost binary classifier for pre-approval social engineering detection: 127 source-combination experiments, segment analysis (OOT Gini 0.841 vs 0.736 full model), RFE + Optuna tuning (935 features → 7) — OOT Gini 0.80, 50% fraud capture at top-5% cutoff. • Built a FastAPI microservice automating fraud analyst report generation (collector pattern, 9 modules, 7+ PostgreSQL sources) with Qwen2.5-VL 35B for AI extraction from complaint PDFs — reduced report time from 25–40 min to 10–15 sec.
