Research & Applied ML

Three public data layers joined to ask whether the AI data-center boom lines up with "100% renewable" messaging and what households actually pay: PJM capacity-auction results, EIA-861 retail residential prices, and EPA eGRID + EIA-923 generation and carbon intensity. Focused on PJM with Northern Virginia as the cluster, with interactive Panel/HoloViews dashboards.

• PJM capacity prices rose sharply after 2024 with high clears projected into 2026–27
• Dominion cleared at a record for 2025/26 and became its own zone on projected demand
• Cleared capacity is still mostly gas, coal, and nuclear

A small CNN trained to tell authentic photographs from spliced (doctored) ones, then stress-tested: accuracy is measured as Gaussian blur and contrast transforms grow, producing degradation curves that show where the detector stops working. Includes explainability passes over the decisions.

Two detectors for AI-generated images built side by side. A random forest over engineered features (statistical moments through fifth order, DCT coefficients, FFT frequency ratios) against a fine-tuned ResNet50, both explained with LIME and Integrated Gradients on a 975-image Kaggle set.

• Random forest ~71%, ResNet50 ~79%
• Variance and the 5th-order moment were the most important engineered features; the ResNet attends to texture and edge patterns

A graph convolutional network that predicts whether a molecule is toxic from its structure, trained on the Tox21 dataset with molecules as graphs and adjacency matrices as the input representation. Two weeks of independent research for a machine-learning seminar, presented to a panel of six professors including the heads of the CS and Math departments, with a one-hour recorded lecture on how GCNs differ from ordinary neural networks.

• ~75% accuracy on Tox21 with a small GCN
• Adjacency-matrix construction and its spectral properties worked out by hand before touching PyTorch Geometric