Hello, I'm
Senior Associate at Capital One, in the Refinance line of business.
I used to engineer biology; now I engineer insights. Five years of turning messy data into dashboards, forecasts and models that people actually use to decide things.
I started in biotechnology, where my dissertation turned out to be mostly about data analysis. That pulled me into analytics, and I have spent the five years since turning messy data into dashboards, forecasts and models that teams use to make decisions.
Today I'm at Capital One, in the Refinance line of business, turning large data ecosystems into decisions that help customers. Before that I was a graduate student assistant at Temple University and a product analytics consultant for Unilever, helping global stakeholders see where their money was going and how to make more of it.
I work across machine learning, statistical modeling and presenting findings to leadership, from sizing untapped markets for a Colgate-Palmolive campaign to cost-optimized fraud models. Outside work I build end-to-end ML projects with real data and an app you can try, several of them from Kaggle competitions.
75% lower fraud cost than no model
A fraud model tuned for money, not accuracy. Only 0.17% of transactions are fraud, so I picked the alert threshold that minimizes review costs plus missed fraud, and added SHAP explanations so an analyst can see why each transaction was flagged.
In progress NFL Big Data Bowl 2027
My entry for the NFL Big Data Bowl: 10 Hz tracking of every Combine drill, matched to the same players' game tracking, to test what the Combine really tells a team about how a rookie will move on Sundays. The write-up and code go public after the January 2027 deadline.
0.9604 AUC, up from 0.9583
Kaggle competition with 700k passengers. Treating flight distance as a route ID, leak-free target encoding and a model trained on the original survey lifted a gradient-boosting blend by 348 leaderboard places.
37% lower error than the baseline
Kaggle competition: predict rock-layer depth along horizontal wells. Physics-constrained path matching instead of a black-box regressor.
1,000 cities routed by road and sea
Chooses where to put distribution centers worldwide using road and sea freight distances, World Bank data, capacity limits and running costs, with an exact solver seeded by a fast local search.
55% win rate on out-of-sample bets
Tests whether Over/Under lines can be beaten with team strength, form, travel and weather. A walk-forward backtest over 5,460 games finds the edge only when betting selectively, and almost all of it on Unders.
On-device model and encrypted memory
A practical take on agentic AI: daily money briefings, anomaly checks and Q&A over Telegram, with the model and my financial history kept on my own machine.
