Large Language Model (LLM) architectures such as DeepSeek-V3 demonstrate reductions in computational costs through the design of efficient architectures, but their development reveals technical debt in the software engineering principles associated with LLM development, partly caused by the multidisciplinary nature of the field. Consequently, it leads to increased development costs and challenges in software quality. To address these issues, this thesis proposes the integration of Model-Driven Engineering into the LLM development life cycle. A conceptual metamodel formalises LLM architectural constructs (RQ1), enabling automated code generation via model transformations (RQ2) and facilitating Model-Based testing (RQ3)