Using data to improve public services in Pittsburgh.

A Pittsburgh-area nonprofit tracked a decade of orders by hand across inconsistent spreadsheets and tallied reports weekly. This is a semester-long IServe Pitt project. Client details are kept general here.

Done

Consolidated 15 spreadsheets (2015–2026) into one clean dataset in R, standardizing names, categories, and dates, and corrected calculation errors found through data validation.

Underway

Rebuilding the Excel tracking with relational design and SUMIFS formulas so weekly reports update automatically, plus an inventory dashboard and data-entry standards for leadership.

A course project that turns a 12,000+ record NHL draft dataset (1963–2022) into a relational database.

Done

Conceptual model with six entities and unique identifiers. The same player name can belong to different people, and one person can be drafted twice, so names alone cannot identify a record.

Next

Logical and physical models, then building and populating the database in SQL.

311 is the number residents use to report non-emergency problems such as potholes and missed trash pickup. I plan to use the City of Pittsburgh's open 311 data from the Western Pennsylvania Regional Data Center to build a few simple measures of how the city responds.

Plan

Define three to five KPIs, such as request volume, time to close, and repeat requests, by category and neighborhood. Publish a short dashboard and a write-up of what I found and what I would track next.

Final project for Intro to Data-Centric Computing at Pitt, August 2026. I built and evaluated six linear regression models that predict a song’s popularity on Spotify, and used cross-validation to confirm the models held up on data they had not seen.

Capstone at Penn State, January–May 2025. I led a team of four building wildfire prediction models with U-Net and LSTM architectures on more than 500GB of satellite data. We built preprocessing pipelines to combine multi-spectral satellite imagery with weather data, and independently validated the methodology of a Stanford study against 12 years of Google Earth Engine data.

Applied data sciences project at Penn State, fall 2024. I built a Bayesian neural network to predict colectomy likelihood from more than 10,000 records. It outperformed deterministic models on edge cases, where knowing how uncertain a prediction is matters most. I added uncertainty quantification so each prediction came with a confidence interval, which is what makes a model usable for decision support.