Businesses that apply AI to mathematical optimization soon hit a wall: models that ace textbook problems collapse on real data. Operations research has long driven routing, scheduling and supply‑chain decisions, yet turning a vague goal into a precise formulation still needs human skill. Recent work shows large language models can generate correct solver code for well‑specified linear and mixed‑integer programs, but demos mask a deeper flaw.

Existing LLM‑for‑OR tools assume a complete, unambiguous prompt with all data embedded. Real supply‑chain projects involve millions of demand rows, multiple products, and constraints that analysts leave implicit, such as facility opening costs or route availability. Feeding raw tables into a prompt inflates hallucinations and exceeds context windows, while the required derived parameters—like distance matrices—must be engineered separately, a step current tools ignore.

ORPilot is an open‑source AI agent built for production‑grade optimization. It begins with an interview stage that asks targeted questions until the objective, variables, constraints and data needs are crystal clear, then a data‑collection stage that defines CSV schemas and validates files, followed by a parameter‑computation stage that derives matrices from raw inputs. The pipeline delivers a solver‑agnostic model ready for reproducible runs.