Open-source platforms such as Hugging Face have become central infrastructure for Artificial Intelligence (AI), hosting large collections of machine learning (ML) models, large language models (LLMs), datasets, and pipelines. These ecosystems enable rapid reuse and experimentation, but little is known about how AI artifacts evolve over time, which factors drive their quality, and how to engineer them in a reliable, maintainable, and sustainable way. This doctoral thesis adopts an empirical Software Engineering (SE) perspective to: (i) characterize the lifecycle and evolution of AI artifacts in large-scale repositories; (ii) identify and model the determinants of key quality attributes, including maintainability, evolvability, and energy efficiency; and (iii) turn these insights into decision-support frameworks and tools for practitioners in AI-based systems. The thesis combines large-scale repository mining, longitudinal analyses, controlled experiments, and tool-building. Overall, it aims to offer an empirically grounded foundation for engineering, evaluating, and supporting quality-aware AI artifacts across their lifecycle.