Yana Savva

• Developed an LSTM-based model for parking lot demand prediction with relative error=8% (Python, PyTorch, BigQuery) • Built and maintained a ML training pipeline (Python, ZenML) • Designed Dagster jobs for model inference (Python, Dagster) • Improved a tree-based revenue predicting model (Python, LightGBM, Scikit Learn)

• Developed time-series transformer for currency rate prediction with sMAPE=0.004 (Python, Hugging Face, PyTorch) • Created NLP pipeline for news sentiment analysis including models with F1 score = 0.95 (Python, Hugging Face, NewsAPI, PyTorch) • Proposed an outlier detection model for stock data using Spectral Residual and CNN architecture with F1 = 0.8 (Python, PyTorch, Scikit Learn, Tensor Board)

• Formed user profile, predicting over 100 key commercial features using exiting user activity data (Jupyter Notebook, Vertica, Lighthouse, PostgreSQL) • Designed recommender system for e-commerce with DAU = 51.1 million (Python, Collaborative filtering)

• Built and improved a model predicting temperature for DAU =10 million (Python, Catboost, DAG) • Created a Computer Vision model for cloudiness segmentation on satellite shots with F1 = 0.92, worked on the associated pipeline (Python, C++, Opencv, DAG) • Refined recommender system of a streaming service with news publications, conducted A/B tests showing 10% higher Click Rate (Java, Python, DAG) • Developed a model predicting the time for push notifications with 20% higher retention (Python, Catboost, SQL)

• Enhanced banking scoring model with 10% higher accuracy (Jupyter Notebook, Python, PostgreSQL) • Engineered a base recommender system for online shop of bank’s debtors’ assets (Python, LightGBM, FastText)