Why AI Still Can’t Solve Your Real Mathematical Optimization Problem

ORPilot is an open-source AI agent designed to handle real-world mathematical optimization problems that existing LLM-based tools fail at. Most LLM-for-OR tools assume complete, small, prompt-embeddable problem descriptions — but real operations research problems are ambiguous, data-heavy, and require derived parameters. ORPilot addresses this through a five-stage sequential pipeline: an interview agent that clarifies the problem before modeling begins, a data collection agent that handles large CSV-based datasets separately from prompts, a parameter computation agent that derives model-ready values from raw data, a code generation agent with sandboxed execution and retry logic supporting five solver backends, and a reporter agent that translates results into plain English. The system was tested on a supply chain problem with 9.7 million decision variables and 963,000 constraints, successfully producing an optimal solution end-to-end.