Yurii

YURII FILIPENKO
Product / Data Analyst — SQL, Experimentation & Statistics
[відкрити контакти](див. вище в блоці «контактна інформація») | [відкрити контакти](див. вище в блоці «контактна інформація») | Paphos, Cyprus | Open to remote
[відкрити контакти](див. вище в блоці «контактна інформація») | github.com/yurii-droid | t.me/Oyroo
SUMMARY
Analyst with a statistics degree working on product and business metrics. Replaced an analyst’s entire manual reporting
workflow with dozens of scheduled pipelines, and built a retention-style model over a 4.2M-row panel (AUC 0.75, 3.3x lift in
the top decile). Strong on A/B design, power analysis and advanced SQL. BSc in Statistics.
EXPERIENCE
Machine Learning Engineer
Prionex LTD | Paphos, Cyprus (on-site) | 2026-Present
Business & product analytics
• Replaced an analyst’s entire manual reporting workflow with dozens of scheduled Python pipelines reading straight from
source databases, retiring a Metabase / Google Sheets / manual-cleanup process; each validated 1:1 against the legacy
formulas.
• Ran the analysis the operations team acts on: campaign targeting and conversion analysis, user segmentation into rule-
based action lists, agent-level performance metrics, and daily revenue reporting reconciled to the cent — querying a
57M-row event table past a 2,000-row API limit with DuckDB as the local SQL engine.
• Measured a 12-month performance baseline across 205 agents and ~2.7M interactions, quantifying a 2–4x spread
between top and bottom performers — the benchmark used to size the opportunity for an intervention.
Modelling & experimentation
• Built a discrete-time survival model — the family used for retention and churn — predicting monthly conversion
probability over a 4.2M-row panel assembled across 11 source databases: AUC 0.75, KS 0.38 on an 839K-row hold-out at
a 1.37% base rate.
• Delivered 3.3x lift in the top decile; the shipped top-25% targeting policy captures 61% of all converting accounts, with a
capacity table letting operations re-set the cutoff against real team headcount. Now running in production.
• Designed a powered champion-challenger experiment — primary metric revenue per agent-hour, minimum detectable
effect 10% at p<0.05 — including sample sizing and group allocation. Separately found a data defect that had mislabelled
~85% of converting accounts and was inflating reported performance from a true 0.75 to 0.91.
Machine Learning Engineer
Confidential (under NDA) | Remote | 2025
• Built and deployed a real-time ensemble scoring engine (rank-correlation stability 0.86, drift monitoring) feeding an
automated decision system.
• Ran 100+ controlled experiments benchmarking competing approaches, establishing the underlying signal was
fundamentally linear; engineered 250+ features validated by walk-forward testing across 45 expanding windows. Earlier,
as analyst, built real-time pipelines over 2.5M+ records feeding Tableau dashboards for stakeholders.
SKILLS
SQL & Data: Advanced SQL (window functions, CTEs, subqueries, query optimisation) across PostgreSQL and DuckDB; large
transactional and event-level datasets; ETL pipelines; data-quality validation; reporting automation
Experimentation & Statistics: A/B and champion-challenger design, hypothesis testing, power analysis and minimum
detectable effect, sample sizing, control groups, statistical significance, sampling, time-series analysis
Product & Business Analytics: Conversion and funnel analysis, user segmentation and targeting, cohort and vintage analysis,
retention and churn modelling (survival analysis), performance benchmarking, campaign analysis, unit-level revenue
reporting
Modelling: Logistic regression, gradient boosting (LightGBM, CatBoost, XGBoost), survival / discrete-time hazard models,
feature engineering, model validation, calibration, gains and lift analysis, uplift-style targeting
Tools: Python (pandas, NumPy, scikit-learn, statsmodels, LightGBM, CatBoost, Matplotlib, Seaborn); Tableau, Metabase,
Google Sheets / Excel; Git
Languages: Ukrainian and Russian (native), English (B2)
SELECTED PROJECT
Credit Risk: Home Credit Default Prediction github.com/yurii-droid/home-credit-default
Optuna-tuned LightGBM classifier, 572 features across 6 tables (0.79 AUC), with temporal validation, isotonic calibration,
cost-aware thresholds and SHAP.
EDUCATION
BSc Statistics, Oles Honchar Dnipro National University | 2021-2025