The rapid adoption of Artificial Intelligence (AI) in the energy sector raises significant challenges related to trustworthiness, regulatory compliance, scalability, and interoperability. While AI-based energy solutions are increasingly deployed across power grids, buildings, and distributed energy systems, there is a lack of unified infrastructures that enable systematic testing, validation, and experimentation under realistic conditions. This paper presents the EnergyGuard architecture, a European AI Testing and Experimentation Facility (TEF) designed to support the development, validation, and acceptance of AI-driven energy solutions. EnergyGuard provides an integrated platform combining secure data management, AI development and experimentation environments, high-performance computing resources, and a dedicated Trustworthy AI acceptance layer. The architecture supports the full AI lifecycle from initial experimentation and trustworthiness assessment to model deployment and continuous monitoring, enabling AI systems to adapt over time in response to evolving operational conditions, data drift and emerging compliance requirements. The architecture is aligned with major European digital initiatives, including data spaces, AI-on-Demand, and emerging regulatory frameworks such as the EU AI Act. Rather than focusing on algorithmic performance, this paper contributes a reference architecture that operationalizes trustworthy-by-design principles and enables scalable, policy-aligned AI experimentation for the energy domain.

Efficient resource management remains a major challenge in machine learning (ML) pipelines, particularly during data preprocessing stages where resource demands are difficult to predict and often lead to inefficient scheduling and resource over-provisioning. This paper introduces AIRE (AI-based Resource Estimation), a framework designed to estimate CPU and memory requirements of individual data preprocessing tasks using dataset characteristics and task-level features. Using an experimental dataset derived from multiple preprocessing tasks and datasets, we train and evaluate several predictive models. Results show that a Random Forest regression model achieves the best performance, reaching up to 99\% accuracy in estimating resource usage. The proposed approach enables more intelligent scheduling of preprocessing tasks, reducing execution time and improving infrastructure utilization. These findings highlight the potential of AI-driven resource estimation to support efficient and scalable management of ML pipelines in modern MLOps environments.

The paper focuses on the oversight mechanisms in relation to the Artificial Intelligence (AI) systems. It explores, including in a comparative fashion, their main advantages and challenges over fully autonomous systems, both from a technical standpoint and from a legal and ethical perspective. The study lingers over how such mechanisms, rotating around the efficient adoption of the human judgment or feedback, can be adopted to address critical gaps in accuracy, accountability and fairness, thereby facilitating the creation of AI systems that learn and improve over time, leading to a performance enhancement and improved model accuracy, also in low-data scenarios. When combined with explainable AI (XAI) techniques, the human oversight paradigm paves the way to build trust and enhance transparency to limit the occurrence of "black-box" models. On the other hand, such a paradigm also might bring issues, including the scalability challenges -due to its impact on costs and speed, which might result in slowing down the decision making process-, the risk of human errors, the difficulties in handling noise in labels provided by human experts, the hurdles in terms of user experience (cognitive load and fatigue for the human experts), the complexities in identifying and adopting the most effective methods for incorporating, storing, and analyzing human feedback in view of improving the model performance. Further challenges include the risk of amplifying, instead of reducing, existing biases in training data, the risk of meaningless human oversight (“rubber stamping”), the difficulties in allocating legal liability for errors in AI systems in case the humans had approved the recommendations made by the AI or when the human operators are blamed even though they had limited ability to override the system. In addition, the risks of privacy violations and data misuse might arise, especially in case the humans have access to sensitive or personal data. Ethical dilemmas also occur, particularly in high-stakes environments like medical diagnosis, in case of unclear boundaries regarding when to trust the AI and when to override it. The concept, advantages and issues are deepened relying at first on a literary review and then with a deep dive on the AI-DAPT environment, highlighting the vision, techniques and measures chosen to enhance the benefits and to prevent or limit the challenges, including also a deep diving on the healthcare domain and its pilot case within the project. Key insights and findings are outlined, as well as the way ahead in this research field

Hybrid models combine first-principles knowledge with machine learning to enhance predictive performance while preserving physical consistency and interpretability. Despite their advantages, such approaches are often developed in a problem-specific manner and lack standardized workflows that support reuse and systematic experimentation. To address this challenge, this work proposes a modular, pipeline-based framework for hybrid model engineering within the Data Analytics and Visualization Environment (DAVE). The proposed Hybrid Model Engineering Engine (HMEE) integrates domain knowledge and machine learning components as configurable operators embedded in directed acyclic graph pipelines, enabling structured experimentation and lifecycle management within a unified platform. As a representative hybrid strategy, residual learning is implemented and demonstrated in a real-world residential energy use case. A physics-informed thermal model provides baseline predictions, while a machine learning model learns to correct systematic deviations through residual modelling. The results illustrate how hybrid workflows can be engineered, executed, and evaluated in a structured and reproducible manner.