Live service that records every arrival prediction the MBTA publishes, observes when trains actually arrive, and learns the bias at each stop to serve a corrected time. Cut arrival error over 25% on a held-out test set.

• Built a Python and SQLite pipeline collecting ~4 million observations per day, halving storage by writing only revised predictions
• Reduced model training input over 13x by pre-aggregating outcome rows into per-stop, per-lead-time error bins
• Trained per-stop, per-lead-time bias corrections across seven rapid transit lines and branches
• Automated hourly processing and daily retraining on EC2 with dead-man's-switch alerting and a public status page
• Built a live Leaflet map, stop lookup, and accuracy page serving corrected arrival times

Logistic regression classifier built from scratch in NumPy to predict NFL regular-season outcomes using quarterback data spanning 80+ years.

• Implemented gradient descent, sigmoid activation, and cross-entropy loss from scratch
• Engineered features from 1940s-2024 quarterback data using Pandas
• Led a 3-person team through data collection, modeling, and code review
• Built interactive visualizations with Matplotlib, Seaborn, and Plotly

Web-based platform for analyzing financial time-series data, computing key performance metrics, and enabling interactive data exploration for investment decisions.

• Built dynamic data pipelines to retrieve, filter, and analyze stock market data
• Implemented data processing logic to compute key asset performance metrics
• Integrated interactive frontend components for data exploration

A local command-line password manager with AES encryption.

• Implemented AES encryption with unique per-vault salts to prevent pre-computed attacks
• Master password is never stored - derived into a runtime encryption key using key derivation
• Built a secure password generator and clipboard integration to avoid on-screen credential exposure
• Designed full CRUD operations via an interactive CLI with password input masking