My research explores how structure—linguistic, musical, or societal—can be modeled, learned, and optimized. I work across foundation model evaluation, robust natural language processing, music signal analysis, and network-based optimization. I'm especially interested in projects that bridge theory and real-world application, using tools from machine learning, optimization, and signal processing to tackle complex, high-impact problems.
Conducting foundation model analysis and determining the effect of AI development on the scientific method and societal impact. Key questions include assessing the accessibility and democratization of AI in science and evaluating the performance of open-source AI models against established data benchmarks.
Constructed an augmented feature set combining MFCCs with wavelet transforms to improve genre classification. Demonstrated that hybrid features boost accuracy for low-capacity models and offer a richer, more robust representation of musical signals.
Enhancing transformer-based sentiment analysis models, such as BERT, by integrating informal language variants into training data to improve robustness against slang, misspellings, and emojis.
Investigating the use of transformer models for video-to-audio sequence generation, optimizing single GPU inference with encoder-decoder architectures across three parameter scales (60M, 150M, 350M).
Developed a mixed-integer program using JuMP and Gurobi to model large-scale hurricane evacuation in Florida. Integrated real-world road, shelter, and population data to minimize unallocated evacuees under capacity and distance constraints. Achieved a 79% improvement over greedy heuristics and identified high-impact infrastructure sites.
Remotely operated a research-grade telescope to capture near-Earth asteroid images and estimate orbital trajectories. Applied astrodynamics and statistical modeling to compute orbital parameters and impact probability.