Projects
Augmented Elo Rating System for Football Match Outcome Prediction
Developed an augmented Elo rating system that learns team-strength adjustments directly from match data, extending the canonical Elo framework with a four-stage optimisation pipeline: backpropagation through time, ISTA for sparse feature selection, SAGA for stochastic optimisation, and cross-validation for model selection. Rating system is compareed against exisiting methods like Dixon-Coles and FIFA World Ranking, as well as market implied probabilites. Built as part of a three-person capstone partnered with the Trilemma Foundation.
Tracking Team Strength: Bayesian Inference for Football League Standings
Modelled evolving team strength across four seasons (2021–25) of FA Women’s Super League data using a Bayesian linear regression fit via Gibbs sampling and a Kalman filter for dynamic strength tracking. The Gibbs sampler achieved an RMSE of 1.440, narrowly outperforming OLS (1.441), while the Kalman filter reached 57.3% match outcome accuracy and correlated strongly (Spearman ~0.90–0.98) with actual league standings.
F1 Miami Grand Prix Lap Time Dataset — SCORE Network
Co-authored a published case study and dataset (1,138 rows of lap-by-lap timing data from the 2023 F1 Miami Grand Prix) for the SCORE Network’s sports data repository, a resource used in statistics classrooms for teaching summary statistics, distribution analysis, and outlier detection. Available publicly with citation.
Kuduk & Ramler (2026) · DOI · Dataset page
Predicting Flight Delay Severity at ORD, JFK & ATL
Framed flight delay prediction as a four-class severity classification problem (no delay, minor, major, severe/cancellation) using five years of US Bureau of Transportation Statistics data covering ~3.5M flights across three major hub airports. Built and evaluated six model configurations in Apache Spark MLlib — logistic regression, decision tree, and KNN baselines alongside bagging, boosting, and a stacking ensemble — comparing both predictive performance and distributed scalability metrics such as training time, speedup ratio, and memory usage.
Predictive Forecasting of Citi Bike Station Demand in NYC
Developed a spatio-temporal graph attention network (MWI-DSTGAT) to forecast hourly bike inflow and outflow at 700+ NYC Citi Bike stations, supporting smarter rebalancing logistics and reducing empty or full-station scenarios. Constructed a spatial graph from station coordinates and flow patterns, incorporated NYC hourly weather data via a dedicated injection mechanism, and benchmarked against LSTM, GRU, and ARIMA baselines across ~30M trip records spanning 2013–2023. Implemented in PyTorch, departing transparently from the assignment’s TensorFlow specification.
F1 Popularity Growth & Sustainability Analysis
Investigated the relationship between F1’s global popularity growth and its sustainability challenges across multiple decades of race calendar data. Integrated circuit location, attendance, and YouTube engagement data from the official F1 channel to explore how calendar expansion, track tenure, and travel intensity have evolved alongside fan demand. Analysed the impact of Drive to Survive on digital engagement trends and examined whether calendar geography has been optimised in line with F1’s net-zero-by-2030 commitment.
Interactive F1 Qualifying Analysis
Built an interactive Shiny application for exploring F1 telemetry and qualifying data across the 2018–2024 seasons. Users can compare two drivers head-to-head within a race or track a single driver’s performance over time through customizable, publication-quality visualizations.
AI-Augmented Physical Therapy App
Developed a cross-platform app for athletes and individuals managing chronic injuries, featuring scheduling, calendar views, and a customizable exercise library. Designed a relational data model for exercises, routines, and user progress, backed by a PostgreSQL database in Docker. Leveraged AI-assisted development workflows (ChatGPT, GitHub Copilot) throughout the build.