The long
version

LoadBrief turns a free-text athlete monitoring summary — training load, heart-rate variability and wellness scores — into a structured load-management brief written for an athlete, a coach or a sports scientist. The fine-tuned model reached 0.960 exact risk-classification accuracy, up from 0.000 for the untuned base model. Auditing that number became the real project.

The vertical cup-to-disc ratio, measured from the optic nerve head, is a key structural marker in glaucoma screening. This capstone, sponsored by UVA Ophthalmology, built a reproducible pipeline that segments the optic disc and cup from retinal fundus images and estimates that ratio, then measured how well a model trained on public data works on real clinic images. It received the Most Innovative Analytical Solution award.

A deep reinforcement learning system for daily stock trading, built on the FinRL framework with PPO agents, VGG and Cross-Stock Transformer feature extractors, and FinBERT sentiment over Polygon news. Single-seed backtests looked strong. Tested across random seeds and stock universes over the full test year, no configuration reliably beat an equal-weight buy-and-hold portfolio.