DESIGN AND DEVELOPMENT OF AN MRS/MLOPS WORKFLOW FOR AI-DRIVEN EXPERIMENTS IN THE SLICES RESEARCH INFRASTRUCTURE

Integrating Artificial Intelligence (AI) and Machine Learning (ML) models into next-generation telecommunication networks (B5G/6G) requires automated, reproducible, and traceable experimental workflows. Developed within the framework of the European research infrastructure SLICES-RI, this thesis addresses the challenges associated with automating the MLOps lifecycle using agents based on Large Language Models (LLMs). Despite the potential of autonomous agents, allowing an LLM to directly and freely invoke execution tools introduces significant issues—such as high latency, non-determinism, and "hallucinations"—that compromise the reliability and scientific reproducibility of experiments. To overcome these limitations, the thesis proposes an architectural paradigm shift: the AI ​​agent no longer acts as a direct executor but rather as a high-level interpreter. Upon receiving a scientific intent expressed by the user (e.g., "train ML models to optimize the energy consumption of a Post-5G xApp and compare their performance"), the agent generates a structured, deterministic execution specification. This specification details infrastructure resources, datasets, dependencies, workflow steps, and expected artifacts. Actual execution is delegated to an external workflow engine (orchestrator) that processes tasks deterministically. In this context, tools based on the Model Context Protocol (MCP) are no longer freely invoked by the LLM; instead, they serve as service interfaces (connectors) strictly controlled by the orchestrator. Concurrently, the role of the Model Registry System (MRS) is radically expanded: it evolves from a simple repository for the final storage of models into a continuous tracking system. The MRS now records the experiment's entire lifecycle, capturing the initial intent, generated specifications, the progress of individual execution steps, data provenance, and every event or error within the cycle. The result is an end-to-end MLOps framework that combines the intuitiveness and flexibility of LLMs during the ideation and planning phases with the rigor, determinism, and traceability (aligned with FAIR data principles) required for advanced scientific research in the mobile networks domain.