my work

Investigating how representational similarity measures — RSA, CKA, and related metrics — capture, or fail to capture, the true geometry of neural and model representations. The project probes how flexible each measure is to benign transformations of representational geometry versus how sensitive it is to meaningful structural differences, to clarify which measures can be trusted for comparing brains and models.

• benchmarking representational similarity measures for sensitivity and invariance across synthetic and real representational geometries
• building diagnostic test cases that reveal when standard measures conflate distinct representational structures

Using representational similarity analysis to compare classical vision models (HMAX, Gabor pyramids) against state of the art deep networks, benchmarked against human fMRI data from early visual cortex. The goal is to understand what structural properties of a model drive brain alignment, and whether simplicity can beat scale.

• found that HMAX exhibits stronger alignment with V1 than recent hyper-complex architectures
• accepted as extended abstract at UniReps, NeurIPS 2025 and contributed talk at VSS 2026

A synthetic image generation pipeline and model introspection framework studying how CNNs handle visual search in structured environments. This URCA grant funded project asked whether neural networks solve visual search the same way humans do, or whether they cheat by relying on statistical regularities in backgrounds rather than the target itself.

• trained multiple CNN architectures to >90% accuracy; found models rely on background context rather than target features
• presented as poster and oral talk at UCSB URCA Week 2025

Studied how humans and large language models evaluate fake expertise in image captions — whether AI systems can detect and be deceived by misleading descriptions. We curated 360 images with intentionally misleading descriptions across five knowledge domains, collected human deceptiveness ratings, and compared them to GPT-4 and Gemini-2.5 judgments.

An AI-powered academic workspace integrating Canvas, Gradescope, and Piazza into a single interface. An agentic AI layer reasons over course context to interpret assignments, surface deadlines, and guide study workflows. FastAPI backend with authenticated APIs and async task orchestration; Next.js frontend deployed on Vercel with Supabase for auth and persistence.

Three frameworks modeling covert attention during dynamic gaze cueing: a Bayesian Ideal Observer, a 3D CNN, and a time-resolved Drift Diffusion Model calibrated against human psychophysics data. Validated against 9 observers and 10,000 trials, replicating key RT speed-up and accuracy effects.

React Native app with GCP NLP sentiment analysis and web scraping, aggregating professor reviews to help students make informed course decisions.

Predicts run, pass, punt, or kickoff from ten seasons of play by play data using a custom Keras model with extensive feature engineering.