RU | EN
Hi, I'm Dmitrii, a Data Scientist / ML Engineer focused on applied machine learning, tabular ML, computer vision, anti-fraud, churn prediction, and business-oriented analytics.
I build end-to-end ML solutions: from data analysis, feature engineering, validation, and model training to interpretation, reproducible pipelines, and business recommendations.
- Tabular Machine Learning: classification, scoring, churn prediction, risk modeling
- Computer Vision: image classification, agricultural monitoring, PyTorch pipelines
- Anti-Fraud: highly imbalanced data, time-based validation, rolling/window features
- NLP / LLM: text features, transformers, Hugging Face, RAG concepts
- Data Analysis: EDA, hypothesis testing, business metrics, reporting
- ML Engineering: reproducible experiments, modular pipelines, model validation
- Churn Prediction: ROC-AUC 0.9045, Accuracy 0.907
- Anti-Fraud / DFC2026: PR-AUC private 0.11144, 200M+ transactions, 0.006% positive class
- Computer Vision / AgroVision: Accuracy 0.896, weighted F1-score 0.897, macro F1-score 0.891
- Business Analytics: reduced slow-moving inventory from 40% to 25%, improved revenue forecast accuracy by 22%
Built an end-to-end ML pipeline for predicting telecom customer churn using SQLite data from multiple tables: contract, personal, internet, and phone.
- Performed EDA, missing value handling, feature engineering, and correlation analysis
- Created business-oriented features such as contract duration and number of services
- Trained and compared Logistic Regression, Decision Tree, Random Forest, CatBoost, LightGBM, and XGBoost
- Best model: XGBoost
- Test metrics: ROC-AUC 0.9045, Accuracy 0.907
- Prepared business recommendations for retention campaigns
Stack: Python, pandas, NumPy, SQLite, scikit-learn, XGBoost, CatBoost, LightGBM, Matplotlib, phik
Developed a computer vision module for classifying agricultural crops from UAV images.
- Built an image processing and model training pipeline
- Trained a PyTorch classification model for 5 crop classes: buckwheat, corn, potato, sunflower, wheat
- Achieved Accuracy 0.896, weighted F1-score 0.897, macro F1-score 0.891
- Analyzed class-level errors and identified quality issues for further dataset improvement
Stack: Python, PyTorch, OpenCV, pandas, NumPy, scikit-learn, Matplotlib
Developed an anti-fraud model for detecting suspicious banking transactions on a large-scale transactional dataset.
- Worked with 200M+ transactions
- Handled extreme class imbalance: 0.006% positive class
- Built time-based validation to avoid data leakage
- Created behavioral, rolling, window, and transaction-level features
- Trained and compared LightGBM, CatBoost, XGBoost, linear models, and GRU-based approaches
- Result: PR-AUC public 0.11101 / private 0.11144
- Private leaderboard: 137 / 574
Stack: Python, pandas, polars, PySpark, LightGBM, CatBoost, XGBoost, PyTorch, GRU, PR-AUC
Built a risk scoring model for predicting at-fault road accidents.
- Performed preprocessing, EDA, feature engineering, and model validation
- Compared CatBoost, LightGBM, Random Forest, and baseline models
- Optimized ROC-AUC and analyzed risk factors
- Prepared business-oriented recommendations for risk segmentation
Stack: Python, pandas, scikit-learn, CatBoost, LightGBM, Random Forest, Matplotlib
Calibrated low-cost electrochemical air quality sensors using machine learning.
- Compared field and winter evaluation scenarios
- Built Random Forest models for NO2 concentration prediction
- Evaluated model quality using MAE and R²
- Analyzed sensor behavior and model stability across conditions
Stack: Python, pandas, scikit-learn, Random Forest, Matplotlib
- Pet-* — applied ML, NLP, CV, and tabular ML projects
- Kaggle-* — competitions and dataset exploration
- Study-* — educational and academic projects
- Case-* — business-oriented analytical and ML cases
Stats are shown for overview purposes only.
Primary focus: project quality, reproducibility, validation, and applied machine learning.
