Skip to main content

Generative AI/Monitoring

GenAI applications in healthcare require new approaches for strong alignment, probabilistic guarantees on hallucination, and adaptive guardrails. My prior work in this area includes:

High Dimensional Data Monitoring
Detecting events from millions of data streams in real-time.

I formalized novel machine learning tasks for real-time data monitoring and developed AI methods for monitoring. These methods were deployed in an applied national public health setting where they led to a 200x increase in monitoring efficiency.

Foundations for Steering Theory
Quantify GenAI model steering capacity.

The interface between interpretability tools and successful long-term model monitoring and guardrailing is underexplored. I developed methods and metrics to quantify model steering capacity when model neurons are monosemantic (encode a single concept) that can be monitored for various adversarial attacks.

Models and Emergent Properties
Human evaluation and distance metrics

Humans are increasingly relying on Generative AI models that can be highly persuasive. These properties are rapidly changing; how do we develop evaluations that keep up?

Selected Papers and Scholarly Work

Software, Writing, and Teaching

Presentations / Collaborators

I’ve presented at or collaborated with MLCommons, MLCommons Rising Stars Program , Computational and Data Science Rising Stars, IJCAI, AAAI, CMU’s Parallel Data Laboratory, CMU’s AI Seminar, CMU’s 3 Minute Thesis competition finalist, AAAI Doctoral Consortium, SIGKDD Doctoral Consortium, University of Ohio, University of Iowa, and the University of Utah. I’ve reviewed for AAAI, IJCAI, CHI, KDD, and VIS.