My research so far sits where machine learning meets its social consequences: measuring bias in generative models, probing how language models fail under adversarial pressure, and studying how public institutions should decide which tasks to hand to automated systems.
Papers & Posters
-
Formalizes a TPR/FPR-based framework for measuring emotion bias in text-to-image models. Across roughly 10,000 prompt–image pairs, finds a consistent and robust bias toward fear in generated images, validated across model architectures and against human annotation.
-
LockdownGPT: Assessing Adversarial Attacks on LLMs
A technical demonstration of systemic vulnerabilities that large language models face across their development lifecycle: jailbreak classes, red-team prompts, and stress tests of common guardrails. Presented to government officials including the former Deputy CTO of the White House Office of Science and Technology Policy.
In Progress
-
Human oversight burdens and AI task allocation in public administration
A conceptual and empirical study of how administrative processes should divide work between officials and automated systems, and what meaningful human oversight costs in practice. I lead the project and am first-authoring the paper with Chris Schmitz and a colleague.
Earlier
-
NLP pipeline for U.S. defense posture analysis
Semantic filtering, LLM-based structured extraction, HDBSCAN clustering, and narrative synthesis over NDAAs, National Security Strategies, and National Defense Strategies, tracking shifts in U.S. force-posture commitments to Europe.
-
Computer vision for autonomous following
A full-pipeline neural-network image classifier and a computer vision program that let robots autonomously follow target items.
Writing
-
A brief on the Commission's argument that the United States must be prepared to deter near-simultaneous challenges from two nuclear adversaries, and what that implies for nuclear strategy, force structure, and the nuclear enterprise.