my work

• built a synthetic dataset generator simulating structured visual environments
• varied background noise & target positions across tens of thousands of images
• trained & fine‑tuned models to 90% human‑level accuracy
• discovered models over‑rely on background cues—potentially insightful for medical‑imaging safety
• pipeline implemented in pytorch

• re-implemented classical vision models (e.g., hmax, gabor pyramids, steerable pyramids)
• ran a curated set of natural & synthetic stimuli through each architecture
• extracted layer-wise activations and compared them to primate neural data
• main finding: hmax's hand-crafted features remain highly brain-aligned
• accepted at unireps 2025!

• studying how humans and llms rate "fake expertise" in everyday image captions
• curated 360 images with intentionally misleading descriptions across 5 domains
• collected human deceptiveness ratings; compared to gpt-4 & gemini-2.5 judgments
• running a follow-up to test decision-making under deceptive information

synthetic image generator + model-introspection toolbox built with pytorch, numpy, and sklearn. comprehensive analysis of how convolutional neural networks process visual search tasks with varying background complexity.

bayesian ideal observer, 3d-cnn, and drift diffusion model that replicate attentional orienting effects. comprehensive framework for understanding how gaze cues influence visual attention in dynamic scenes.

react native mobile application with gcp nlp sentiment analysis and web‑scraping capabilities using beautifulsoup. helps students make informed decisions about course selection based on aggregated professor reviews and ratings.

predicts run/pass/punt/kickoff using ten seasons of play-by-play data with a custom keras model. comprehensive feature engineering and model optimization for accurate real-time play prediction.