Facilitated interactions between 2,000 Marsh employees and clients by developing an interactive Power BI dashboard that analyzes insurance pricing trends across 9,715 firms over 10 yearsIdentified a negative correlation between claim frequency and years since federal judge appointment (R = -0.143) by scraping 2,000 federal judges and fuzzy matching on Advisen loss dataImproved root mean squared error (RMSE) of Directors and Officers claim frequency model by 0.4% by training, tuning, and comparing 12 XGBoost modelsReduced spreadsheet processing time by 15 man-hours through developing Python scripts using openpyxl to automate Excel tasks to combine Advisen loss data with S&P Capital IQ data

Launched a project with Dr. Zhaozheng Yin training GANs with spatial transformer networks to generate handwritten wordsHarnessed IAM and CVL datasets to evaluate GAN performance with different writers in English and GermanBuilt scripts to test GANs qualitatively through image generation and quantitatively through FID Score

Detected gerrymandering in 5 states by applying the Polsby-Popper test through a React.js interface, leading to the formulation and dispatch of district plan proposals to relevant state authoritiesConceptualized 15 district plans using recombination Markov chains on 2020 House of Representatives districts, establishing baselines for comparative analysisExpanded available data by 100% by leading a team of 3 to clean 2020 precinct-level election result shapefiles from MIT Election Lab using GeoPandas, managing data through Amazon RDS, and testing changes with Spring Boot

Generated 50 peer groups from 7.6 million workers compensation claims by writing SQL queries on Apache Hive and experimenting with dimensionality reduction algorithms such as UMAP and PacMAPRemoved 10% of claims with irrelevant or missing information by analyzing accident descriptions through word clouds and count plots made using Seaborn