A benchmark for evaluating the ethical alignment of LLM agents when moral imperatives conflict with profit incentives. It embeds the prisoner's dilemma and public goods game in morally charged contexts, varying moral framing, opponent behaviour and survival pressure across nine models, and estimates each factor's causal effect while analysing motives in agents' reasoning traces. No model remains consistently moral; game structure and moral framing are the strongest drivers, exposing the situational brittleness of current models.

Evaluating a feature-attribution method usually means deleting the tokens it deems unimportant, which pushes the input off-distribution and throws away the nuance in the scores. We instead replace those tokens with plausible substitutes found through Gemma embeddings and faiss similarity search, so an explanation reflects the attribution scores it came from.

Research on bringing AI capabilities into Oracle Database. Designed and built a multi-agent LLM application that automates financial investigations (patent submitted), and contribute to WayFlow and Agent Spec, Oracle's open-source frameworks for executing and specifying LLM agents.

Built and deployed a model for automated processing of transaction confirmation PDFs, and prototyped an entity resolution solution matching bolted joints across different production manuals.

Worked independently across plant production, machine operation and livestock, and was involved in planning and implementing renewable energy projects (wind, solar, biogas).

Coursework in Probabilistic AI, Machine Perception, Design of Parallel and High-Performance Computing, Interactive Machine Learning, and Reliable and Trustworthy AI.

Coursework in Deep Learning, Operations Research and Optimization, and Data Analysis & Simulation. Spent an exchange term at the University of the Fraser Valley in Abbotsford, Canada.

A pipeline for instance segmentation of individual trees in Rwandan satellite imagery, comparing Mask R-CNN and Cascade Mask R-CNN in PyTorch and MMDetection, from tiling and annotation handling through to evaluation on held-out regions.

Implemented and tuned kernels for a set of PolyBench-style benchmarks in both CUDA C and CUDA Julia, then measured where Julia's abstractions cost performance on GPU and where they don't.

A React and Spark application for comparing Graph Neural Network explainability methods on a molecule toxicology dataset. Domain experts can interact with the molecule graph, reassign edge explanation weights by hand, and watch the resulting metrics recompute on the fly.