About Me

I recently completed a postdoctoral appointment as a visiting assistant professor at the University of Notre Dame after completing my Ph.D. in mathematics from Northwestern University in 2018 under the direction of Ben Weinkove and Valentino Tosatti.

Current projects in data science. Master's coursework in statistics, neural networks, time series, data mining, and machine learning methods. Experience in Python (NumPy, pandas, scikit-learn, Keras), R, Mathematica, Excel, MATLAB.

Previous mathematics research on nonlinear PDEs in geometric analysis and complex geometry: Analysis of singular elliptic and parabolic complex Monge-Ampère equations, Chern-Ricci flow, conical Kähler metrics, and Calabi-Yau geometry. My Erdos number is 6.

Resume Department Website LinkedIn Github

Projects

Optimization of Bid Strategy for Online Ad Rankings: Root Insurance Challenge

Designed a bidding strategy using anonymized advertising data provided by Root Insurance to optimize cost per customer acquisition for online advertisements. Identified customers likely to purchase policies, modeled competing advertisers' bids, and implemented stochastic gradient descent and constrained optimization in Python to determine cost effective bid sizes for different customers. Presented results to peers and industry experts as part of the Erdős Institute data science boot camp, scoring near the top among all projects. (Github) (Drive)

Deep Neural Collaborative Filtering Recommender for Comics and Graphic Novels

Used the UCSD Book Graph dataset containing user data from Goodreads.com to build collaborative filtering recommendations for comics and graphic novels. Fit matrix factorization and deep neural network models in Keras to produce recommendations which returned a Top-10 Hit Rate of 0.93 on validation data. (Article) (Github)

Data Mining and Statistical Learning: Final Project

The final project for data mining and statistical learning was to use provided data to build a classification model to predict the class of the target variable using 100 feature columns. The instructions were to return the predictions of two chosen models, one logistic regression model and one non-GLM family model. The model predictions are evaluated with final score based on ROC-AUC score of test predictions. I used several different models in each family to fit the training data, and used 10-fold cross-validation to estimate mean test ROC-AUC score, determining the best model for each family. The models ultimately used to generate training predictions were logistic regression with recursive feature elimination, and AdaBoost with a small learning rate. (Github)