• Completed a multiclass computer vision model (resnet50, seresnext101,etc) on aerial imagery to detect varying levels of damage by peril after a large loss event. The project also involves creating a pipeline to mask and preprocess the images for instant scoring in production *Built NLP models (CNN with embeddings, LSTM with embeddings, tfidf) to detect potential P&C regulatory complaints within approximately 8M member messages in the Claims Communication Center (CCOM) transcripts. Invented a new way to remove words from transcripts, reducing processing time from over 24 hours to ten minutes. Innovated to build a robust pipeline to clean data of PII, misspelled words, and unusual words using NER, customized algorithms, autocorrect models, embeddings, and complex regex rules. Sharing methodology to team to be incorporated into future NLP projects. • Created a water peril territory model (lightgbm) using geospatial attributes, Bayesian hyperparameter tuning, and new data sampling and aggregation schemes to improve lift, reduce overfitting and model drift, and expedite modeling time. Initial results showed a 22% improvement in loss cost lift and 73% improvement in loss ratio gini. • Invented in a geospatial clustering algorithm that groups properties into statistically credible neighborhoods based on latitude and longitude of the property. USAA filed patent PC-2515.00_171-1584 in March 2022. • Built a highly accurate model (lightgbm) to predict the probability of a non-weather water (nww) claim within the policy year for use in loss prevention, underwriting, marketing, pricing, and segmentation. Pulled and preprocessed over 40M unique HO policies to create a database of in-force policies with over 300 modeling attributes. The model segments both high and low risk policies with a result of over 32 times spread in loss cost between the riskiest 2.5% of policies and lowest risk 2.5% policies. • Won first place in the P&C Hackathon 2022

• Built models (gbm, svm, double clustered knn, dbscan, birch, gmm) to segment customers and predict customer churn for each product category for commercial customers of all sizes, industries, and locations. The model accuracy was 96%. Cleaned and aggregated over 95 million rows of purchase order data and mapped locations to each customer for geolocation accuracy. Feature engineered over 1,500 attributes to capture frequency, recency, trend, seasonality, and market basket behavior of customers and added over 300 census data attributes and 200 trended IRS attributes for modeling and segmentation. • Created a series of intermittent demand models (deconstructed hurdle, croston) to forecast Amazon purchase order demand by distribution center for GP commercial package goods sold through Amazon. Using limited data, reverse engineered Amazon’s supply chain to calculate pseudo inventory levels of GP products at Amazon’s fulfillment centers. Combined the supply chain results with Amazon ecommerce data, GP purchase orders history, and brick and mortar point of sale data to create a final modeling dataset with time series data and exogenous attributes. The models were able to forecast orders locations and timing almost 90% of the time within an acceptable margin of error, resulting in a meeting with Amazon’s VP of consumer packaged goods and a renegotiation of the unfilled order penalties. • Constructed churn and customer lifetime value models (gbm, svm, linear regression) for the commercial corrugated business, allowing account managers to optimize customer contact and product offerings for new and existing businesses. • Managed and mentored a team of data scientists for consumer-packaged goods projects across Georgia Pacific

• Utilized machine learning techniques and advanced statistics to model for space elasticity curves (sales and margin as a function of space) at a macro (category) and micro (SKU) level for over 600 store product categories. The models also include an automated feature selection macro that better accounts for multicollinearity and streamlines the inclusion of many additional modeling attributes• Wrote SQL and Python scripts to gather, clean, and merge many additional data sources for inclusion in the optimization models and reports. Feature engineered over 1300 new modeling attributes for the space optimization models. • Created algorithms to clean and correct historical store data, which erroneously reported store category space ¼ of the time. Aggregated, deseasonalized, and detrended corrected store data to measure lift of the space elasticity curves against historical space/sales changes in stores. • Supported the field (store managers) by completing ad hoc projects for Home Depot stores and providing analytics to help maximize KPI growth

• Utilized machine learning techniques and other feature selection algorithms to create the Attract 5.0 countrywide credit model for underwriting auto insurance. The model was developed on approximately 6.5 million countrywide policies and 5,500 of the latest enhanced credit attributes, including new trended credit attributes, and it was controlled for common insurance rating attributes.• Wrote a macro for supervised weight-of-evidence binnings on numerical variables given a continuous target, such as loss ratio or loss cost. The result is significantly expedited modeling EDA and variable preprocessing. The previous solution, smbinning, utilized a binary target, which did not match the model targets. • Built the InsurView models using a large dataset of 46 million countrywide auto policies and public record data in order to help insurers underwrite auto insurance policies. The InsurView models are Tweedie pseudo loss ratio model, and contain a control score that accounts for the common rating variables of age, gender, territory, credit, and property ownership. The result is that 89% of auto policies that are thin-file/no-hit customers (using credit) can be scored and 55% of customers receive a better segmented rate. • Co-wrote and presented the property insurance trends report, Home Trends 2017 and Home Trends 2018, as well as several research papers published on PC 360 and LexisNexis’ website.

• Wrote SAS code/macros and documentation to import, clean, and analyze premium and loss data for 136 profitability study business segments for the Actual versus Expected (AvE) template to be used on all commercial lines.• Revised rating algorithms and restructured all financial lines raters to restore functionality and accurately export policy details to SharePoint. Also updated the Excel macro exportation macros, SharePoint list settings, and Access extractor tools in order to migrate all of Financial Lines global business to a new SharePoint site• Used VBA and Excel to create and automate the rate monitoring reports, account quality index reports, and profitability studies

Using C to create statistical models. Will later apply these models to C Cuda.Used R to optimize portfolios of catastrophe bonds given risk limits, portfolio size, spreads, expected loss, and increment size constraints. Compared outcomes of various optimizations by modeling the portfolio as a an approximated discrete, finite case and a continuous case and computed the VaR and CVaR for the optimizations.

• Completed an 18-month rotational training program in civil site, roadway, and traffic design• Prepared construction plans, cost estimates, project reports and schedules, calculations, and models for the various roadway, civil site, and traffic projects• Designed the site layout, grading and drainage, utility, erosion control, and construction detail plans for eight Navy Federal Credit Union buildings across the United States.