Research

A single principle runs through our research: modular reasoning, the ability to understand and change one part of a system without grappling with the whole. We design the languages and interfaces that make modular reasoning possible, and we reframe software-analysis tasks so that this reasoning scales to millions of programs. Today the principle matters most for AI-enabled systems: we decompose learned models into modules and bring software-engineering rigor to the AI that society increasingly depends on, so it can be built, debugged, and trusted.

View our work on Google Scholar →

Complete publication record: Google Scholar, DBLP, and our lab.

We design languages, interfaces, and contracts that let engineers reason about complex software one module at a time, from aspect-oriented design and Classpects to Ptolemy's quantified, typed events, capsule-oriented Panini, and translucid contracts for modular verification.

The hard question behind Boa was not how to store hundreds of thousands of repositories, but how to express software-analysis tasks modularly so they can be decomposed and run across a massive collection of programs. Boa is the language and infrastructure that answers it, and the result is large-scale empirical software engineering the whole community can use.

Modular reasoning extended to AI-enabled systems. We decompose learned models such as deep neural networks into reusable, replaceable modules, and we work to make AI software dependable through the first comprehensive study of deep-learning bugs, fault localization and repair for neural networks, and fairness across machine-learning pipelines.