Ravin Kabra

• Maximizing hearing aid optimization by designing a cascading ensemble model (XGBoost + Transformers), achieving 97.5% accuracy and 96.5% ROC AUC using A100 GPUs with distributed computing. • Strengthening classification accuracy by applying transfer learning on Google’s YAMNet combined with a cascading ensemble model, reaching 94% accuracy and 91% F1 score. • Ensuring ethical data use by curating and analyzing open-source audio datasets while maintaining license compliance.

• Reduced downstream appeal costs by $250K+ annually by deploying a CatBoost model to predict claim appeals at intake using claims data. • Implemented a time series anomaly detection system that combines STL decomposition with Isolation Forest to detect claim irregularities. • Accelerated appeal triage by 58% by developing a HIPAA-compliant LLaMA-4 Maverick and Claude Opus 4 agent on Databricks using PyTesseract, rule-based tagging, and few-shot prompting to classify scanned appeals and verify single-member content.

• Reviewing research papers and applying current findings and methodologies to project needs. • Collaborating with the PI to develop and apply data science and machine learning concepts (gradient boosting, PCA, ICA, neural networks, etc.) to refine STRF models. • Building and optimizing backwards and forwards models from scratch using MATLAB, focusing on time series analysis, neural coding, signal processing, data preprocessing, and feature engineering. • Developing MATLAB programs for STRF estimation using advanced signal processing techniques like Fourier Transform and spectrogram computation. • Conducting EEG readings using Biosemi and EMOTIV, and evaluating domain-specific bootstrapped statistics and metrics. • Performing feature engineering and time series analysis to create time-lagged features, visualizing STRF matrices and auditory data, and analyzing the impact of parameters on spectrogram accuracy and model performance.

• Improved Kroger’s retail staffing decisions by developing a CatBoost model to predict optimal associate staffing levels per store and week, reducing forecast errors by 92%, and cutting training time by 91% through parallel computing and automation. • Supported HR and leadership in compensation decisions by building interactive dashboards with SQL Server, Power BI, and ArcGIS to visualize pay equity, geographic trends, and tenure-performance patterns. • Supported employee retention strategy by clustering 350,000+ employees using DBSCAN (silhouette score: 0.71) to identify high-performing personas at risk of churn based on tenure, performance, and pay metrics. • Investigated drivers of Full-Time % (FT%) across stores using OLS regression and causal inference techniques on variables, including staffing levels, store type, geography, and department mix; identified significant predictors influencing FT composition. • Conducted in-depth interviews (IDIs) with store managers, HR business partners, and labor planning leads to gather qualitative insights on operational challenges, FT/PT mix preferences, and local labor constraints; triangulated findings with statistical models. • Integrated quantitative and qualitative insights to validate regression findings and highlight actionable levers such as cross-training hours, union presence, and regional norms in FT hiring. • Developed visual summaries of IDI themes and matched them to regression coefficients to support leadership in refining labor strategies and addressing staffing equity across divisions.