Research and projects

Producer prices are supposed to lead consumer prices. Testing that against 78 years of U.S. data turns up a more interesting answer than the headline number suggests. A model combining lagged PPI with lagged CPI predicts inflation with an R² near 0.97, but almost all of that accuracy comes from the persistence of inflation itself. Measured against a CPI-only baseline, what PPI actually adds is about 2%, and that gain cannot be distinguished from zero.

• Modeled 78 years of monthly U.S. CPI and PPI data, from 1947 to 2025, on a chronological 2010 to 2025 hold-out period
• Built ridge regression models across 1 to 12 month lags, selecting lag length by BIC on training data alone
• Benchmarked against a CPI-only autoregressive baseline to isolate what PPI adds beyond inflation’s own persistence
• Found the 2% RMSE gain from PPI statistically indistinguishable from zero under Diebold-Mariano testing
• Validated with expanding-window refits and Newey-West standard errors across 733 training observations
• Measured a 26% rise in forecast error during the 2020 to 2021 COVID-19 disruption

A chatbot for contact lens manufacturing operations that lets plant supervisors, maintenance engineers, and quality teams query production data in plain English instead of waiting on a data team.

• Designed a star schema across 12 tables in a Databricks lakehouse, modeling production events and error incidents against dimensions for plants, machines, shifts, products, and error types
• Built an intelligence layer combining clustering to group recurring errors by pattern, z-score anomaly detection to flag deviations in temperature, yield, and cycle time, and correlation analysis to surface co-occurring errors
• Configured a Genie AI interface that resolves natural-language questions into SQL across the full schema and returns results with visualizations
• Stress-tested the system at 100× dataset scale, identifying how statistical assumptions in the anomaly layer degraded accuracy on complex root-cause queries
• Presented the architecture and a live demo to company leadership

An interpretable text classifier built from scratch, then benchmarked against a standard machine learning baseline to see how far transparent, hand-engineered features can get you.

• Split 3,309 labeled sentences into 2,647 training and 662 test examples
• Engineered word-frequency, sentence-length, and punctuation features using log-odds band analysis
• Scored 69.18% with the rule-based classifier against a 70.24% TF-IDF and kNN benchmark

A complete real-time game written in Python and Pygame, presented as a final project and live demonstration at Tufts.

• Implemented a game loop handling input, player movement, enemy waves, and rendering
• Used object-oriented architecture to manage players, enemies, and game state

A 12-week online data science course taught by Dr. Robert Barton, an Amazon machine learning scientist. Covered data interpretation and visualization in Python and Jupyter Notebooks, building and validating mathematical models with statistical analysis, and a final project analyzing text sentiment.

A summer software development programme, finishing with a formal technical presentation and live demo to Tufts faculty, peers, and parents.