EPAM Systems
Senior Data Scientist
Client: Opella (Sanofi Consumer Healthcare). Weekly e-commerce sell-out forecasting for Amazon and other retailers in 8+ markets, on Snowflake, SageMaker Pipelines and MLflow with Chronos-2, Prophet, MFLES and Ridge models.
- Product discontinuation lifecycle in the production forecaster: phased-out products fade to zero, are capped by retailer stock and hand volume over to their successor.
- Model vs. covariate study over 27 rolling backtest origins: showed model choice is worth 6–9 WAPE points while external data adds under 1, and ensembling gains about 2.
- Plackett-Burman factorial screen of feature groups in 16 parallel SageMaker runs, replacing one-feature-at-a-time testing; cut screening time by 63% (246 → 91 min) and SHAP cost by about 60%.
- Redesigned cannibalization detection to remove false positives from zero-inflated sales; found by reproducing runs that production had silently switched to the wrong Chronos checkpoint.
- Audited the 9 forecast-exclusion rules (12.7% of recent volume excluded, several rules dead) and proposed a three-tier redesign; ranked cross-market data-quality issues by revenue impact.
- Self-initiated internal tools: nx, a Snowflake data-access and run-reproduction library used in 40+ notebooks; benchpress, a FastAPI app for benchmarking forecast runs; and ops, a Rust CLI tying Jira, GitHub and SageMaker test runs together.

