Sergei M.

Data Scientist

Belgrade, Serbia10 yrs 3 mos experience

Highly Stable

Key Highlights

Built ML systems for 200M+ users.
Reduced loan default rates by 70%.
Improved moderation accuracy from 50% to 90%.

Stackforce AI infers this person is a Data Scientist specializing in AdTech and FinTech with extensive experience in applied machine learning.

Contact

Skills

Core Skills

Applied Machine LearningLlm SystemsTime Series Forecasting

Other Skills

Vector Search (HNSW)CatBoostBERTRoBERTaA/B TestingLogRegRandom ForestCLIPHNSWETLReal-Time InferenceGitPythonSQLRanking & Retrieval Systems

About

Senior Data Scientist with 5+ years of experience building and deploying production-grade ML systems at scale for 200M+ users in AdTech and FinTech. I specialize in applied machine learning: recommender systems, ranking and retrieval, Trust & Safety / content moderation, credit risk, and time-series forecasting. I work end-to-end across the ML lifecycle - from problem formulation and feature engineering to A/B-tested production impact. Hands-on experience includes large-scale data processing (YTsaurus, Spark), embedding-based retrieval (HNSW), real-time inference, and models based on Transformers, CLIP, and CatBoost. I also build LLM-powered systems with controlled prompting and RAG-style architectures, focusing on reliability and interpretability.

Experience

10 yrs 3 mos

Total Experience

3 yrs 5 mos

Average Tenure

Current Experience

Dzen platform

Senior Data Scientist

Nov 2023 – Mar 2026 · 2 yrs 4 mos · Belgrade · Remote

Fine-tuned transformer-based NLP models (BERT / RoBERTa) with a CatBoost meta-learner for content classification (F1 0.90+), optimized for CPU-only production ML.
Developed an LLM-based moderation service using structured XML prompts for deterministic classification; expanded rare-category coverage by ~50% with A/B-validated lift.
Built and deployed machine learning models (LogReg, Random Forest, CatBoost) for fraud/scam detection, leveraging feature engineering and threshold tuning to eliminate 33% of manual checks.
Built a clustering and multimodal similarity pipeline (CLIP + HNSW) integrated with a vector-search retrieval layer, improving duplicate detection accuracy by 21% and enabling low-latency inference.
Improved data quality via automated annotation and ETL pipelines, boosting moderation model accuracy.
Scaled automated moderation from ~50% to 90%, reducing manual workload by 40% across 200M+ daily users.
Designed and implemented ranking components and retrieval-based signals for content integrity pipelines, improving precision and minimizing inference-time latency in real-time audit flows.
Led full-cycle experimentation: A/B testing, statistical power analysis, and experiment monitoring.

Vector Search (HNSW)Applied Machine LearningLLM Systems

Exiar

Data Scientist

Dec 2021 – Nov 2023 · 1 yr 11 mos · Moscow, Russia

Built a full ML model development and deployment pipeline for a credit risk scoring system (CatBoost), reducing loan default rates by 70% through calibration and optimized decision thresholds.
Developed a revenue prediction time-series forecasting pipeline (CatBoost) that improved accuracy from 13% to 4% WAPE.
Developed a churn prediction model (AUC 0.90+), enabling targeted retention and reducing churn by 15%.
Automated ETL and real-time dashboards, reducing reporting time by 22% and improving anomaly detection.
Identified a customer segment that doubled LTV, increased retention by 36%, and raised conversion by 28%.
Optimized scoring-service latency using lightweight real-time inference techniques for credit decisioning.
Contributed to portfolio growth by integrating machine learning into operational decisioning.

Applied Machine LearningReal-Time InferenceTime Series Forecasting