Shayne Goldstein

Why this matters: Embedding models are foundational to modern AI systems. This research directly addresses AI safety and fairness by identifying and quantifying their biases.

• Analyzed 6 major embedding models across 11 dimensions of potential bias (source credibility, false moral equivalency, framing, and more)
• Found 67% of models demonstrate statistically problematic bias; identified critical failures distinguishing terrorist propaganda from democratic military sources
• Faculty-endorsed as "the core of a good research paper" by Elon CS professor; currently under review for publication

Why this matters: Current methods require physical tagging. This project enables non-invasive, individual-level behavioral studies bridging computer vision and behavioral ecology in meaningful ways.

• Designing a non-invasive camera tunnel system to capture individual salmon in natural stream conditions
• Training a CV re-identification model on existing long-term monitoring datasets, built to handle turbidity, motion blur, and inconsistent lighting
• Deploying on local hardware for cost-effective, human-free field use; full open-source system + deployment instructions as deliverables
• Enables individual-level behavioral studies previously impossible without physical tagging