research

Evaluating ML-KEM at (much) higher security levels

Designing and analyzing a novel noise distribution for ML-KEM at elevated security levels. Currently implementing formally-verifiable, constant-time AVX2 assembly in Jasmin and benchmarking on multiple x86-64 hardware.

Monogenic Fields of Cryptographic Size

Explicit number-field constructions via polynomial transformations for Ring-LWE schemes. Designed and analyzed algebraic number fields in SageMath, computing discriminants, Galois groups, and ramification, to identify cryptographically secure parameters, with proofs of irreducibility and monogenicity. Paper currently in final stages of preparation.

Developed the tolerant, a novel invariant that extends and strengthens classical discriminant theory. Proved a closed-form formula via discriminants and resultants, resolved an open question in prior work, and established its structural and algebraic properties.

A multi-agent LLM system that generates personalized, curriculum-aligned weekly homework by integrating vector-store retrieval, goal planning, and automated grading. Evaluated via an IRB-approved teacher focus group.

Informational Sufficiency in LLMs

A framework measuring when a model has "enough" information to reason well, extending the notion of underspecification. Building a benchmark to detect insufficient inputs, analyze model stability, and identify the minimum information needed for reliable LLM reasoning.