Advancing Formal Privacy and Empirical Privacy Auditing

• Investigated how to best scale user-level differentially private training to LLMs under fixed compute budgets. SaTML 2025 paper.
• Improved privacy auditing methods for LLMs to better detect data leakage and inform practical understanding of model risk. Proposed a simple and effective canary design that leverages high-temperature sampling to yield highly sensitive synthetic probes for detecting privacy leakage. Presented and evaluated an improved approach for auditing synthetically generated data. Investigated memorization properties of language models using these privacy-auditing techniques. Under review at EMNLP 2026.

Models that Forget Innovation Pod

• Led a research effort leveraging modularity and unlearning to control data influence and address privacy and policy challenges. 1 of 4 proposals of bold new research directions selected from 56 submissions across the organization. Built team of collaborators, drew connections across orgs, charted out research roadmap, secured buy-in from leadership.
• Proposed data compartmentalization as a strategy for AI Governance, presenting a taxonomy mapping governance needs to technical requirements and relevant strategies. Presented at GenLaw@ICML 2024.
• Documented internal use cases for unlearning and scoped out partnerships through discussions with product teams.
• Hosted 3 student researchers, investigating approximate unlearning for robustness and certifiable unlearning for privacy. ICLR 2026 paper.
• Contributed to an internal effort on modular continual learning.

Federated Learning

• Published an interactive blog post on federated learning. Through interactive visualizations the piece allows readers to wrestle with data privacy concerns, learning how FL aims to mitigate them while introducing trade-offs.
• Developed infrastructure to support scaling up federated learning simulation to datasets and models orders of magnitudes larger than previously used. Released an open source library, Dataset Grouper, for partitioning existing flat datasets into group-structured datasets. NeurIPS Benchmarks & Datasets 2023 paper.
• Led a student researcher project to explore new algorithms for merging models in federated learning, drawing upon ideas in meta-learning. TMLR 2024 paper.

Compression for Federated Learning

• Designed a custom compression method for client updates in federated learning to reduce the communication cost from 32 bits per model parameter to 0.1 bits without degrading accuracy on several benchmarks.
• Implemented compression-based aggregation methods for federated learning in TensorFlow Federated. This involved: custom TF ops hosted in TensorFlow Compression, TensorFlow Federated logic, and robust integration tests for system compatibility.
• Presented this work to internal and external audiences. Selected to give an oral presentation at the Google Research Conference. Oral presentation at ICLR 2022 TinyPapers.
• Integrated custom compression method into Google’s federated learning production system.

Drug Metabolism Prediction Using Graph-based Learning

• Built a deep graph convolutional network (GCN) using Pytorch to predict drug metabolism.
• Proposed the use of molecular representations learned through GCNs to identify metabolically labile atoms. Compared to traditional feature extraction methods.
• Completed written thesis and oral defense in April 2020.

Developing Civic Scientist Leaders Program

• One of ten graduate students selected to participate in a weekly seminar at the Baker Institute for Public Policy to learn about the federal policymaking process and develop critical leadership skills to advance science as a public good.

Anomaly Detection on Time-Series Metrics

• Developed and implemented an anomaly detection system to automatically detect regressions in quality of service among subsets of our network and generate reports to alert iCloud engineers. Deployed system surfaces one to two critical issues each day that otherwise went unnoticed.
• Presented work to ~30 engineers at iCloud and individually to the Vice President of iCloud.

Benchmarking an Incremental Docking Protocol

• Improved an incremental docking protocol (DINC) which computationally predicts how peptides bind to protein receptors. Experimented to identify unexpected behavior; strengthened the robustness of DINC by handling these edge cases.
• Evaluated the latest version of DINC by designing re-docking experiments and writing scripts to automate these tests on the XSEDE Comet Supercomputer. BMC Molecular and Cell Biology 2019 paper.