Research
Algorithms now function as public infrastructure. My work treats them the way engineers treat bridges: design them to carry load safely, inspect them regularly, and stress-test them against the forces they will actually face.
1. Algorithms with guarantees
When an algorithm allocates resources, clusters people, or aggregates votes, its design encodes a notion of fairness. I design approximation and online algorithms that provably balance fairness against efficiency, and quantify the price of that tradeoff.
Fair hierarchical clustering (NeurIPS 2023, ICML 2023)
Fair and efficient online matching (NeurIPS 2024; under review NeurIPS 2026)
Fair division with entitlements (AAAI 2024) and online max-min allocation (NeurIPS 2022)
Metric distortion in voting (ICLR 2026)
2. Measuring and verifying AI systems
Large language models are becoming decision-making infrastructure, but their safety properties are fragile and hard to certify. I study how alignment breaks, how to attack it efficiently, and how to verify it statistically.
The geometry of alignment collapse under fine-tuning (under review)
Semantically coherent adversarial attacks via diffusion guidance (ICML 2026)
Statistical verification of alignment constraints (EAAMO 2026, oral)
Validity of LLM-based resume screening (IASEAI 2026)
Evaluating AI research capability: TCS-Bench and agentic autoformalization in Lean (under review)
3. AI and public infrastructure
As AI capabilities grow, so does the set of people able to exploit weaknesses in public systems. I examine concrete systems, starting with elections, and work with policymakers to prepare for them.
AI-enabled attacks on ballot secrecy in Georgia (DEF CON 34; paper under review with IEEE SaTML 2027; AP coverage)
Single-transferable-vote modeling for the California Legislature (Data & Democracy Lab)
UN Independent International Scientific Panel on AI, Preliminary Report