PhD candidate in Computer Science, University of Vermont
MassMutual Research Fellow, Computational Ethics Lab
I work on data privacy and machine learning: differentially private data releases, privacy attacks that test them, and LLM pipelines for policy and public-sector documents.
ACM FAccT 2026
Across 2,912 survey responses from 782 college students, comfort sharing personal data depends on who is asking. Discomfort with platforms predicts private settings, and students from marginalized groups or with adverse childhood experiences are less comfortable with institutions of power.
JAMIA Open, 2026
Releasing survey and Oura ring data from 600+ first-year students in the LEMURS study as differentially private synthetic data. AIM, MST, and PATECTGAN are compared across privacy budgets from ε = 1 to 100; AIM at ε = 5 keeps the data useful while resisting privacy attacks.
EPJ Data Science, 2024
A tree of 120+ conspiracy theories, built from the literature and 1,769 fact-checking articles, and a pipeline that finds and sorts conspiracy theories in new text: a RoBERTa classifier (87% F1), UMAP and HDBSCAN clustering, cluster labels, and named entities.
Preprint, 2026. Under review at Government Information Quarterly
Neural OCR, LLM classification, and network modeling for 21,000+ Iowa 28E intergovernmental agreements, with 96% extraction accuracy and 82% weighted F1. A hierarchical stochastic block model of the statewide contract network shows a stratified buyer–seller structure across government types.
Submitted to the Journal of Privacy and Confidentiality
The Multivariate Gaussian Attack models correlated counts across overlapping marginals instead of collapsing them to a single score. On Adult, ACS, and 2020 Census data it reaches an AUC of 0.98 on unprotected small groups, and the privacy budgets of recent Census-style releases (ρ ≥ 20) can still leave small groups exposed.
Computational Ethics Lab
Scores privacy policies against a 40-item GDPR and CCPA framework. Each chunk of a policy updates a Bayesian belief that a requirement is missing, vague, or clear, using embeddings and NLI models, until the system is sure enough. Only uncertain cases go to an LLM. A hand-labeled set of 100 policies is used for training and evaluation.
June 2026, OpenDP Differential Privacy for Health and Genomics Workshop
Harvard University, Boston, MA
Utility trade-offs of differentially private synthetic data generators for LEMURS, a behavioral health study at UVM.
LinkedIn post
June 2026, Theory and Practice of Differential Privacy Workshop
Northeastern University, Boston, MA
A poster on a new membership inference attack for tabular data that fits the attack model with a multivariate Gaussian distribution.
LinkedIn post
May 2025, Governor’s Institutes of Vermont
University of Vermont, Burlington, VT
Data exploration with Python for talented high school students from across Vermont, staying on the UVM campus with the Governor’s Institutes of Vermont math program.
LinkedIn postI’m a computer science PhD candidate at the University of Vermont and a MassMutual Research Fellow in the Computational Ethics Lab. My work sits where differential privacy, machine learning, and data governance meet: releasing sensitive data that people can still use, testing whether releases leak, and reading policies at scale with language models.
I earned my MS in computer science at UVM with a thesis on a computational, genealogical approach to conspiracy theories, and a BS in electrical engineering from Shiraz University. In 2026 I was an R&D intern at MassMutual, working with longitudinal health data from the Living Well app.
Outside research, I serve on the board of Local Motion, advancing safe streets and active transportation across Vermont, and I spend a lot of time on a bike.