Using AI to help non-experts quickly fix broken optimization models

Real-world optimization models often need rapid updates when constraints change or new rules emerge, but modifying them typically requires specialized expertise. This work deploys an LLM as an interactive OR assistant that translates user prompts into structured model modifications, selects appropriate re-optimization techniques from a toolbox, and returns feasible solutions. The toolbox accelerates solving using primal information—historical solutions, valid inequalities, solver tuning—while maintaining solution quality. Tested on supply chain re-optimization (prioritizing speed and proximity to current plans) and university exam scheduling (prioritizing quality), the approach significantly improves computational efficiency and keeps model changes interpretable and traceable.