Middle Data Scientist

Mitigated recency bias in nearline recommendations by rebuilding the training dataset, switching to a next-item objective, and introducing a multi-target anti-bias loss for candidate generation. Removed RAM bottlenecks by moving to an iterable dataset pipeline. Result: GMV +13% and diversity +1.7% in production A/B testing. Improved recommendation diversity by replacing quota-based post-processing with category-aware retrieval built on multiple FAISS indices split by category, then tuning the diversity–relevance trade-off through five A/B iterations. Result: +15.7% diversity and +1.5–5% on key engagement metrics, followed by production launch. Improved search ranking by sourcing and integrating a broad set of item-level features into the ARGUS ranking model and running systematic ablation studies to measure their impact on NDCG and ROC. Built a cold-start media recommendation baseline by replacing manual curation with clustered SigLIP embeddings used to drive exploration for new users. Result: ATPU +7.2% in A/B testing and production deployment. Automated h2h A/B experiments by turning a manual multi-team handoff into a reproducible workflow that scores the correct user cohort and publishes daily recommendations by experiment ID. Stack: Python, PyTorch, PySpark, FAISS, Airflow, S3, MLflow, Triton.