Interpretability · 2026
Seven cyclic task families. 42 directed transfer pairs in Qwen2.5-7B. Matching the donor’s signed operation improved causal transfer despite low neuron overlap.
Manuscript & poster
Research
Looking inside the model.
I want to understand how models use their internal representations to reach decisions, then use that understanding to build more reliable systems.
01
Experience
2025-present
Stanford Causality in Cognition Lab
Mechanistic interpretability researchWorking with Matthew Shinkle on how language models bind people, relations, and objects. I use interchange interventions to test which internal representations affect their answers.
2024-present
SafeBeat Rx
Machine learning research internBuilding clinical tools with physicians and UCSF collaborators, from automated medication extraction to clinical summaries. My current research focuses on predicting biological heart age from ECG signals.
Summer 2026
UCLA COSMOS
Language model circuit researchInvestigated whether circuits with little neuron overlap can perform compatible computations across tasks with shared representational geometry.
02
Selected Work
Language models · 2026
A controlled study of classification-head design with fixed encoder representations, isolating aggregation, attention, and lexical augmentation.
ICLR Sci4DL
Efficient ML · 2026
How communication budgets constrain federated learning, including the cost of transmitting models from the server to participating clients.
IEEE ICFEC
On-device ML · 2026
A MobileNetV3-based system for offline waste classification on consumer devices, with 97.74% validation accuracy across 12 categories.
Healthcare · 2026
Automation to extract medication information from electronic medical records and reduce manual processing in clinical workflows.
Journal of the American College of Cardiology
Language models · 2025
A 3,000-example lexical analogy benchmark spanning synonymy, antonymy, and derivational morphology. Nine models evaluated alongside a 300-example human baseline.
Preprint
Healthcare · 2025
Built the automated social-media data collection infrastructure for a study included in the 30th Annual Atrial Fibrillation Symposium abstracts.
Journal of Cardiovascular Electrophysiology
Peer review
Reviewer for the 2026 NeurIPS Symmetry and Geometry in Neural Representations workshop.