Machine Learning Research · PyTorch, ResNet-18, Linear Probes, Interpretability
Studied how L1, L2, and dropout regularization reshape where and when spurious versus core features are encoded across network depth and training time in a ResNet-18. Using linear probes at six layers on Colored MNIST and Waterbirds, we found that each regularization method constrains spurious features through a qualitatively different mechanism.
Product Management · Product Strategy, Launch Planning, Competitive Research, Product Requirements & MVP Scoping
Authored a product + business plan for Hot Ticket, a fashion-first social platform that connects outfit inspiration to shopping through native brand tagging. Defined market framing, hypotheses, success metrics, and a phased go-to-market/monetization strategy with clear guardrails to preserve authenticity and trust.
Data Science · K-Means Clustering, Jupyter, Statistical Testing, Data Analysis
I worked with egyptologist and anthropologist Dr. Stuart Tyson Smith on a project surrounding Egyptian burial practices. Using data he collected, I analyzed 134 New Kingdom tombs (located in Thebes, Egypt) to identify socioeconomic and cultural patterns in funerary practices.
LLM Training and Development · PyTorch, Transformers, Fine-tuning, Benchmarking
As a research assistant in the Qin Lab at UCSB, I explored methods to improve safety in vision-language models against multimodal jailbreak attacks. I implemented a GAN-style discriminator to align projected image tokens with language tokens, hypothesizing this would transfer the LLM's text-based safety guardrails to image inputs.