Андрій

Andrii Korostashov
Junior ML Engineer / Data Scientist
Ukraine | [open contact info](look above in the "contact info" section)
GitHub: github.com/AndryMilki
LinkedIn: [open contact info](look above in the "contact info" section)

Profile
Junior ML Engineer with a background in Computer Engineering and ongoing studies in Applied Mathematics.
Portfolio experience in classification, time series forecasting, ML inference services, and SQL analytics. Builds
reproducible Python pipelines, evaluates models against baselines, and develops PostgreSQL data models and
Power BI reports.

Technical Skills
Programming and data: Python, SQL, pandas, NumPy, Git, Bash basics
Machine learning: scikit-learn, XGBoost, TensorFlow/Keras, statsmodels, imbalanced-learn; classification,
ARIMA/SARIMAX, cross-validation, threshold tuning, SHAP
Analytics and BI: PostgreSQL, CTEs, window functions, dimensional modeling, Power BI, DAX, Power Query, cohort
retention, RFM segmentation, matplotlib, seaborn
Engineering: FastAPI, REST APIs, Docker, Docker Compose, GitHub Actions CI, unit testing, joblib

Selected Projects
E-commerce Analytics with PostgreSQL and Power BI
Designed a six-table PostgreSQL database with keys, constraints, and indexes using synthetic data covering
2,000 customers, 10,799 orders, and 60,000 sessions.
Created analytical views and SQL analyses using CTEs, LAG, DENSE_RANK, and NTILE for monthly growth,
product rankings, RFM segmentation, cohort retention, and session conversion.
Built a Power BI project with Power Query imports, three fact tables, shared dimensions, DAX measures, and
two report pages for executive KPIs and product performance.
Added SQL assertions for fact grain, payment reconciliation, and cross-fact revenue; aggregated items and
payments before order-level joins to prevent double-counting.
Customer Churn Classification
github.com/AndryMilki/classification-models
Built a churn prediction pipeline comparing Logistic Regression, Random Forest, Gradient Boosting, and
XGBoost, with SMOTE and dummy/no-SMOTE baselines.
Tuned hyperparameters with GridSearchCV and selected a decision threshold on validation data; achieved F1 =
0.595 and PR AUC = 0.566 on a separate test set.
Added a FastAPI prediction endpoint, Docker packaging, tests, CI, and SHAP/feature-importance reports.

Additional Projects
Industrial Production Forecasting with ARIMA and SARIMAX
github.com/AndryMilki/forecast-ARIMA-model
Built a reproducible forecasting pipeline for the U.S. Industrial Production Index with stationarity testing,
ACF/PACF analysis, and residual diagnostics.
Selected ARIMA(4,1,0) using information criteria and compared forecasts with naive and moving-average
baselines using RMSE, MAE, MAPE, and Theil’s U2.
Documented that simple baselines outperformed ARIMA on the selected test split; added future forecasts,
saved reports, unit tests, and CI.
Website Defect Analyzer
github.com/AndryMilki/CNN_defector
Built a Dockerized CNN/Keras inference service for website screenshot classification, with FastAPI endpoi nts for
URL analysis, image prediction, and health checks.
Implemented Selenium screenshot capture and automated website quality checks; added Docker Compose,
Makefile commands, environment configuration, and tests.
Documented inference and deployment scope and model limitations, including the unavailability of the original
training dataset.

Education
V. N. Karazin Kharkiv National University
M.Sc. in Applied Mathematics | 2025–present
Double Degree Program with the University of L’Aquila, Italy | 2025–2026
National Technical University “Kharkiv Polytechnic Institute”
B.Sc. in Computer Engineering | 2021–2025

Languages
English — Upper-Intermediate | Ukrainian — Native