AI-Accelerated Research in Operations

The relationship between artificial intelligence and optimization research has advanced rapidly, and in this talk I describe some stages of my own work in this space as an Amazon Scholar. I begin by discussing a model behind a real application at Amazon in last-mile delivery operations, a non-convex quadratic integer program that proved intractable at scale. Collaborating with a frontier reasoning model, we (AI and myself) uncovered a structural insight that reformulated the problem entirely, with significant consequences for what became solvable in practice. I then ask how capable AI models truly are at graduate-level mathematical modeling. I treated leading AI models as students in a doctoral optimization course that I teach at the University of Washington, grading their solutions against a decade of real student performance to see how they measure up. Recognizing the limits of any single model, I then turn to what becomes possible when many agents work together. I describe a progression of increasingly capable multi-agent designs. Together, these projects trace how AI has reshaped the way I approach optimization research.

Frontier AI models are starting to do something slightly unconventional for theoreticians: make real progress on research-level (not just competition-level) mathematical problems. In this talk, I will describe my journey through an ongoing project, motivated at least in part by self-preservation: getting ready before the models come for me. With my team at Chicago Booth, I have been building an agentic pipeline for attacking open problems in theoretical operations research. The pipeline extracts open problems from papers, formalizes and encodes them, tracks partial progress in the literature, and verifies they are not solved. It then uses AI models to propose and refine mathematical solutions through multiple iterations and collaboration between a network of several LLM-based agents. I will share what has worked, what has failed in amusing ways, and what this suggests about the near future of AI-assisted theory. Time permitting, I will also discuss our plans for democratizing the pipeline through our website and building a shared memory system for better tracking partial progress.

I’ll show how I use computer agents such as Claude Code/Cowork to help with writing papers and prove complex results. If time permits, I’ll also talk about how to use knowledge graphs to supercharge the AI-assisted research process.

A working researcher's field guide to integrating agentic AI into daily practice — covering hardware and software setup, high-leverage workflows, and the practical guardrails that keep the human firmly in charge.