Kagan Agras

Kagan Agras

Hi! I’m Kagan. I am a sophomore studying Computer Science and Cognitive Science at UC Berkeley. I focus on AI interpretability, representation learning, and adversarial robustness.

My most recent work is on model introspection and scalable oversight, building on Transluce’s Predictive Concept Decoders (PCDs) line of work. I independently lead a team of 4 Berkeley undergrads replicating and extending PCDs to decode normalized activations from multiple layers, as final project for Will Fithian’s AI Interpretability course. Our extension achieves 3% improvement in attribute prediction accuracy over the single-layer baseline, with consistent gains on automated interpretability scoring. (blog post coming soon!)

Previously, I worked as a Software Engineer at RescueSight. I built a multimodal RAG pipeline to enable rapid disaster response planning. In high school, I worked on detecting fake news in low resource languages, and presented my work at Regeneron ISEF.

In my free time, I like playing tennis, listening to podcasts, and practicing martial arts. Feel free to reach out if you want to chat!