LLM-assisted cloud migration toward sovereign architectures

Abstract

Cloud-native applications are architecturally coupled to provider-specific managed services, creating vendor lock-in that conflicts with European digital sovereignty requirements. While cloud migration is well-studied, migration away from cloud services toward sovereign, cloud-agnostic architectures remains largely unaddressed, and large language models (LLMs) have not been applied to this problem with deterministic verification guarantees. This thesis asks how an LLM-assisted migration pipeline can be designed to systematically perform such migrations. Following Design Science Research, it presents a four-stage pipeline (Analyse, Test, Transform, Verify) with three verification gates that separates stochastic generation from deterministic verification. In two case studies, the migrated systems pass all three gates, including all specification-driven tests. A 40-run reliability evaluation reveals a prompt complexity paradox: minimal prompts outperform detailed ones, suggesting that LLM training-data knowledge is more reliable than explicit instructions. A seven-theme sovereignty rubric validated with ten industry engineers achieves 91.7% pairwise agreement. The central finding is that reliable LLM-assisted software engineering depends on designing verification infrastructure around the LLM, not on improving the LLM itself.