RDE Labs: Vision and Role

The AIS Lab advances AI security through rigorous, applied research in threat detection, vulnerability assessment, and protective measures. We focus on securing modern AI systems against realistic, adaptive adversaries, bridging the gap between research and production. Our team conducts foundational research in AI security and translates breakthroughs into practical, deployable solutions that integrate seamlessly with your existing stack. We rigorously evaluate models, agents, and AI systems under real-world attack scenarios, uncovering failure modes across the model, data, infrastructure, and application layers. Beyond assessment, we help organizations secure AI releases end to end. This includes designing adaptive guardrails, implementing continuous monitoring, and establishing robust validation and assurance processes that evolve alongside emerging threats.

Detect. Prevent. Comply.
We equip organizations with the tools and knowledge to deploy AI systems safely and responsibly. Detect risk before it becomes an incident. We run controlled adversarial simulations to stress-test AI systems before real attackers can exploit them. This involves simulating malicious behavior through prompt manipulation, tool-chain disruptions, and protocol deviations to uncover vulnerabilities such as jailbreaks, data leaks, unsafe tool activations, and injection flaws. Translate findings into integration checks. You get replayable evidence, a clear fix plan with ownership, and a regression suite integrated with CI so fixes hold and issues do not return.

Block unsafe AI behavior in real time without adding latency.
We enforce real-time input and output policy, keeping prompts, memory, and tool actions within safe bounds while meeting latency targets. Production telemetry feeds new attack patterns and drift signals into red teaming and assurance so defenses improve with every release.

Meet regulations with evidence.
We run evidence‑based evaluations across model behavior, guard prompts, retrieval layers, tool orchestration, and data handling to catch latent vulnerabilities and misalignment early. We align evaluations to regulatory controls (EU AI Act, NIST AI RMF, ISO 42001, IEC 23894), generate signed, audit‑ready traces, and enforce fail‑fast promotion gates.

Simone Gallivanone

Dr. Galllivanone holds a PhD in Mathematics, with a background in differential geometry and microlocal analysis. His research interests include adversarial machine learning, with a focus on applying mathematical tools to its geometric and robustness-related aspects. Outside of research, he enjoys sports and outdoor activities.

Marco Russo

Marco Russo is a PhD candidate in Computer Engineering at Politecnico di Torino, where he recently submitted his dissertation on quantum computing and quantum machine learning. In 2023, he spent two months as a research scientist intern at Fermilab in Chicago, USA, working on the application of machine learning to quantum control of superconducting qubits on low-latency cryogenic electronics, aiming to achieve high-fidelity quantum gates.

Raffaele Paolino

By drastically reducing discovery and development cycles from months or years to days or weeks — and by enabling unprecedented levels of customization — AMED seeks to accelerate the transition from concept to deployment. Its work spans applications ranging from ultrasound devices and adaptive robotic materials to scalable, automated design pipelines for industrial use.

Through this dual focus on foundational advances and translational impact, the lab contributes to sectors including aerospace, automotive, biomedicine, manufacturing, robotics, defense, consumer electronics, and other domains where advanced materials and intelligent engineering design can be transformative.

Rodolfo Subert 

Rodolfo Subert is a Postdoctoral Researcher in the AMED Lab at AI4I, with expertise in computational materials physics and soft matter. Prior to joining the lab, he worked in Prof. Marjolein Dijkstra’s group at Utrecht University, where he studied liquid crystals and self-assembling systems, investigating how microscopic interactions give rise to complex collective behaviors. His background combines statistical mechanics, simulation, and data-driven modelling of structured materials. At AI4I, Rodolfo works on the computational modelling of complex functional material systems and on AI-enabled approaches to the design of robotic matter. His research is particularly relevant to the study of emergent behaviors, integrated sensing and actuation functionalities, and the use of machine learning to uncover new physical mechanisms for advanced materials with capabilities approaching, and potentially surpassing, those found in biological systems.

Marco Benedetti

Marco Benedetti is a Postdoctoral Researcher in the AMED Lab at AI4I, with expertise in physics, machine learning, and complex systems. He obtained his PhD in Physics from Sapienza University of Rome in the group of Nobel Laureate Prof. Giorgio Parisi, where his research focused on associative memory in artificial and biological neural networks. His work brings together statistical physics and learning theory to understand how intelligent systems encode, retain, and generalize information. At AI4I, Marco investigates the foundations of physics-informed diffusion models for scientific and engineering applications. His research focuses in particular on how physical priors can be embedded into generative models to substantially reduce data needs while improving reliability and generalization. Through this work, he contributes to the development of principled AI methods for advanced materials and engineering design.

Binzhou Zuo 

Paolo Secchi 

The AI-based Software Quality Assurance (AQUA) Lab develops next-generation software testing and debugging by combining advances in artificial intelligence with software engineering research. Led by Dr. Alessio Gambi, the lab aims to improve the quality, reliability, trustworthiness, and safety of software-intensive and AI-enabled systems while keeping developers at the center of the quality assurance process. AQUA pursues this vision through three interconnected research streams: AI-assisted and automated test generation and debugging, continuous improvement and optimization of regression testing, and AI-aware software testing education.

By combining AI, search-based algorithms, runtime analysis, and data-driven methods, AQUA develops approaches that support more effective and efficient software development activities, including automated test generation, fault localization, bug fixing, test optimization and management. Alongside research on software testing and debugging, AQUA works towards preparing current and future generations of developers for the ongoing AI-driven transformation of software development.

The AQUA Lab embraces a holistic view of software quality assurance that spans the entire software lifecycle and cuts across application domains wherever software and AI play a critical role, including dependable software systems, cloud and distributed computing, autonomous cyber-physical and robotics systems. Through its unique combination of foundational research, industrial collaboration, and education, AQUA seeks to foster a new culture of responsible AI-driven software quality assurance with broad technological and societal impact.

Core focuses on the lab’s research agenda include: (1) encrypted inference protocols, for running machine learning models while keeping both the model and the queries private, (2) model integrity protocols, that certify the correctness of ML inference and (3) semantic watermarking, for detecting AI-generated content.

The lab’s work spans theory, protocol design, cryptanalysis, and implementation. It is grounded in foundations, seeking a better understanding of the fundamental questions that arise at the cryptography/ML interface, aiming to provide trust solutions with mathematically provable guarantees. At the same time, the research is motivated by practice, with an emphasis on designing practical protocols that can be potentially deployed in the real world.

The lab develops next-generation solutions in Edge AI, wearable and distributed sensing, robotics, smart infrastructure, healthcare technologies, industrial monitoring, and environmental sensing. By integrating innovations in hardware, software, and machine learning, EIS aims to create intelligent systems capable of delivering reliable, real-time decision making while minimizing energy consumption and computational resources.

Through the scientific leadership of Prof. Magno, the lab benefits from a strong international research network, extensive experience in leading and coordinating European research and innovation projects, and long-standing collaborations with world-leading sensing, semiconductor, and technology companies. These connections foster the rapid transfer of cutting-edge research into impactful applications while creating opportunities for collaborative innovation across academia, industry, and society.

The EIS Lab serves as a bridge between scientific excellence and real-world deployment, accelerating the adoption of intelligent sensing and Edge AI technologies across sectors including healthcare, manufacturing, robotics, smart cities, agriculture, transportation, and humanitarian applications.

The Intelligent Development and Large-Scale Systems (IDeaLS) Lab operates within AI4I – Artificial Intelligence for Industries, the national AI center founded by the Italian Government to perform transformative, application-oriented research in Artificial Intelligence. AI4I’s mission is to foster innovation, industrial transformation, and economic growth through cutting-edge AI research and development. With a strong institutional commitment and long-term national funding, AI4I is rapidly evolving into Italy’s leading AI center, shaping the AI research and innovation agenda in Italy and Europe. It builds on a state-of-the-art on-premise HPC cluster, the Leonardo supercomputer at CINECA, and strategic partnerships with the Italian Institute of Technology (IIT), as well as a robust scientific, industrial, and financial ecosystem.

Within this context, the IDeaLS Lab focuses on advancing the foundations and applications of intelligent, adaptive, and dependable software systems.

Its mission is to research and engineer novel methods, tools, and pipelines for the development, testing, and evolution of AI-enabled and cyber-physical systems, combining expertise across Software Engineering (SE), Data Science (DS), Life Engineering (LE), and Cloud Computing (CC).

Research in the IDeaLS Lab explores how AI, Machine Learning, and Generative approaches can enhance software lifecycle activities—from optimized AI training, automated testing, continuous delivery (DevOps/MLOps), runtime monitoring, to code review, maintenance, and intelligent software evolution.
Particular attention is devoted to dependability, explainability, and sustainability of large-scale intelligent systems, ensuring that new AI-driven software ecosystems are transparent, trustworthy, and energy-efficient. The lab promotes interdisciplinary collaborations bridging Software Engineering, Artificial Intelligence, and domain-specific fields such as life sciences, smart cities, robotics, and intelligent infrastructures.

Ongoing and future projects aim to deliver adaptive and resilient software architectures for autonomous and large-scale systems, leveraging empirical methods, cloud-native platforms, and HPC-based experimentation. The IDeaLS Lab aims to serve as a collaborative hub for researchers, industry, and institutions, contributing to AI4I’s overarching vision of developing innovative, sustainable, and impactful AI technologies that empower the next generation of intelligent systems.

Geometric Reinforcement Learning for Lightspeed Adaptation:
By formulating RL problems through the lens of differential geometry, we aim at leveraging the structural biases arising from the policy representation, the underlying dynamical system, and the learning algorithm itself to design resource-efficient RL frameworks. These RL frameworks will unlock fast and efficient adaptation and fine-tuning of robot foundation models, visuomotor policies, among others.

PRAXIS Laboratory develops Verifiable Intelligence for agentic AI: systems that learn from data, coordinate with other systems, and act under explicit guarantees. The lab brings together statistical machine learning, formal methods, and agentic systems to address a central problem in modern AI: how to build systems that adapt to complex operational contexts while making their reliability and admissibility explicit before their outputs affect real processes.

PRAXIS focuses on AI systems that are no longer limited to prediction. In industrial and organizational settings, AI is increasingly expected to interpret procedures, inspect documents, use software tools, and support decisions inside workflows. The obstacle is not only that current models may hallucinate or make mistakes. The deeper issue is that their behavior is not organized around the constraints of the process in which they operate. A useful system must connect learned evidence to admissible action: it must know when the data support a judgment, when a constraint permits the corresponding action, and when uncertainty or ambiguity requires escalation.

The lab’s approach is to design learning, verification, and orchestration as parts of the same system. Statistical learning provides the basis for adaptation from limited and heterogeneous data. PRAXIS develops methods that exploit structure across tasks, outputs, and environments to improve generalization, reduce sample complexity, and quantify uncertainty. This builds on a research trajectory in structured prediction, transfer and meta-learning, optimal transport, and operator world models, with the common aim of understanding when data-driven systems can be trusted beyond the examples on which they were trained.

Formal methods provide the basis for admissible execution. PRAXIS studies specifications in which symbolic constraints coexist with learned predicates. Some requirements can be represented directly, such as budget limits, approval hierarchies, or safety invariants. Others depend on context and must be learned from data and practice. The scientific challenge is to let these learned judgments enter a formal decision process without treating them as unquestionable facts. PRAXIS develops calibrated and verifiable interfaces through which uncertainty can determine whether an action proceeds, is refined, or returns to human control.

This perspective is especially important for agentic and multi-agent systems. When several agents plan, negotiate, or call tools in the same workflow, reliability cannot be reduced to the competence of each component in isolation. The composition of their actions must remain within the intended behavior of the whole process. PRAXIS studies coordination methods in which agents can adapt while remaining accountable to global constraints, and in which disagreement or verification failure becomes information for specification refinement rather than a reason to proceed on informal confidence.

The lab is foundational in its methods and industrial in its orientation. Its collaborations begin from serious deployment problems where AI must operate inside real procedures, not only perform on isolated benchmarks. Industrial partners provide operational boundaries, data regimes, and failure modes; PRAXIS turns them into mathematical questions, algorithms, benchmarks, and software artifacts. The goal is to create research outputs that are scientifically meaningful and practically usable in settings where adaptation and guarantees must coexist.

The name PRAXIS reflects the lab’s aim to keep theory and deployment in close contact: theory provides the discipline needed for reliable action, while deployment reveals which assumptions, guarantees, and abstractions matter in practice.

Under construction

_

Under construction

_

Under construction

_

Under construction

_

Under construction

_