
Modernising business-critical legacy systems with language models usually means sending proprietary code to cloud services, which raises intellectual property and security concerns that block adoption in industry. This paper reports on modernising a terminal operating system used at several sites of a major European port operator: over 22,000 source files, about 10,000 of them C#, a multi-layered web forms interface, a complex Oracle database and extensive documentation. Local models extract a formal four-view specification, covering domain models, behavioural rules, process state machines and screens, across 48 bounded contexts, so the source never leaves the premises. Cloud models then build new implementations from the specifications alone. On a CRM probe the best local model reached a structural fidelity of 0.83, while two of seven candidates could not produce parseable JSON. Prompt engineering raised the semantic purity of the domain model from 0.71 to 0.94; old documentation added almost nothing. Working test-first, the team generated 5,451 behavioural tests before any implementation and produced two runnable stacks, in Python/FastAPI and .NET 8 Blazor.
Authors: Andreas Biesdorf, Eckhard Seibert, Nicole Wengatz, Bart Carpels, Tom Schoonbaert
Contributing partner: Trier University of Applied Sciences