OptiChat: Making Optimization Models Explainable with Agentic AI (in collaboration with Chewy)
OptiChat: This project explores how agentic AI, optimization, and data analysis can be combined to make large-scale decision models more interpretable in real-world settings. In collaboration with Chewy, students will help extend the OptiChat framework, which uses large language models and structured tool use to explain optimization results, compare model runs, interpret sensitivity information, and draw insights from historical solutions without requiring repeated re-solves. The project is especially relevant for students interested in optimization, machine learning, and practical decision-support systems, and offers hands-on experience at the intersection of optimization, AI, and industry applications.
Mentors:
Description:
Optimization models are mathematical tools used to make the best possible decisions under given constraints, such as determining how much inventory to order, how to schedule deliveries, or how to allocate limited resources in a supply chain, and they are widely used in industry, including at companies like Chewy where large-scale linear and mixed-integer programs are solved on a daily or weekly basis to support critical operations. Despite their power, a major challenge for practitioners is that the outputs of these models are often difficult to interpret, especially when inputs change over time. For example, when two runs of the same model produce different solutions, it is not always clear what caused the change; similarly, while tools like sensitivity analysis provide information such as shadow prices, these quantities are often abstract and hard to translate into actionable insights, and in many real-world settings re-solving large models to explore “what-if” scenarios is computationally expensive or impractical. This VIP project, in collaboration with Chewy, builds on our OptiChat framework, an agentic AI system that leverages large language models together with structured function calls to interact directly with optimization models, solution data, and analytical tools, enabling the system not only to generate explanations but also to actively query results, compare runs, and perform targeted analyses in a tool-augmented manner. The project focuses on three key questions motivated by Chewy’s real-world needs: how to explain differences between solutions obtained from different input datasets, how to interpret sensitivity information to predict how solutions will change under small perturbations, and how to leverage historical optimization runs to explain current decisions without requiring additional solves. By combining optimization, data, and LLM-driven reasoning in an interactive framework, this project aims to bridge the gap between mathematical models and human understanding, enabling practitioners to better trust and act on optimization outputs while providing undergraduate students with hands-on experience at the intersection of operations research, machine learning, and real industrial applications
Students will learn how to connect optimization, data analysis, and agentic AI to build interpretable decision-support tools in real-world settings. We are particularly looking for motivated undergraduate students in computer science, industrial engineering, or related fields who are interested in optimization, machine learning, and building practical systems.
Relevant Technologies:
- Optimization connection
- Data analysis
- Agentic AI
- Optimization
- Machine learning
Pre-requisite knowledge/skills:
- We are particularly looking for motivated undergraduate students in computer science, industrial engineering, or related fields who are interested in optimization, machine learning, and building practical systems.