Recommending Energy Retrofits at the Neighborhood Level

We develop computational methods for identifying and prioritizing residential energy-efficiency retrofits at neighborhood scale. This work combines building-energy simulation, synthetic housing data, machine learning, and heterogeneous evidence sources to model retrofit opportunities when comprehensive building-level data or ground-truth labels are unavailable. Methodological challenges arising from this work have also motivated broader research on constructing supervision from heterogeneous evidence without ground truth.

Collaborators: Jorge Silveyra, Chetan Tiwari

Trustworthiness of Surrogate Explanations for Neural Networks

I study when surrogate models provide trustworthy explanations of neural network behavior. This work examines the relationship between predictive accuracy, surrogate fidelity, and functional agreement, from global measures such as the λ-score to the spatial structure of teacher–surrogate disagreement. More broadly, the project investigates when apparently faithful surrogate explanations genuinely recover model behavior, where they fail, and how those failures can be characterized.

Honors Thesis: Computational Characterization of Idiomatic Language

My honors thesis investigates how linguistic theories of idiomaticity can be translated into computational representations for natural language processing. Focusing on properties such as semantic opacity, syntactic flexibility, and conventionality, I am developing methods for characterizing idiomatic expressions and evaluating whether learned language representations capture the distinctions predicted by linguistic theory.

Advisors: Sofia Serrano

Aller de soi : les idioms face au modèle saussurien [Taking it for Granted: Idioms verses the Saussureian Model]

This independent study in the French program examines idiomatic language through the lens of modern linguistic theory. Focusing on French idioms, I analyze the limitations of Saussure's structural model in explaining idiomatic meaning and explore how more recent linguistic frameworks provide a foundation for future computational models of idiomaticity.

Advisors: Maria Hernandez