As AI becomes increasingly integrated into critical sectors, ensuring robustness, transparency, and regulatory compliance is essential to harness its potential responsibly and effectively. Industrial, healthcare, and financial applications face growing demands for explainable, resilient, and trustworthy systems. Meeting these challenges requires approaches that combine technical innovation with adherence to evolving standards and regulations, enabling AI to deliver real-world impact safely and reliably.
Objectives
The Trustworthy AI Group combines cutting-edge inter- and multi-disciplinary research with industrial-grade MLOps and compliance-oriented design to deliver transparent, resilient, and regulation-aligned AI solutions. Its work addresses a broad spectrum of real-world applications, including high-risk, high-impact solutions in sectors such as healthcare, finance, defence or deep-space observation, adhering to Luxembourg’s national AI strategy and the EU AI Act’s risk-based governance framework. More specifically, the group aims to co-designing trustworthy AI systems through rigorous benchmarking, explainability and strong alignment to regulatory knowledge across domains.
Research focuses on:
- High-impact and high-risk applied AI:
Activities that connect AI with healthcare or projects such as MILAN2, for deep space image processing, which stem from a previous collaboration with the French company Vaonis. The group enabled the development of 100+ custom AI models, leading to 2 patents and a large, curated reference dataset. The project also produced licensable solutions for mobile-embedded multi-frame super-resolution astrophotography and environmental monitoring segment, demonstrating real-time, on-device AI prowess.
Another example is the 3D-4Land defence project (EDA, Q1 2025), which focuses on situational awareness and 3D landscape representation using multimodal sensors (LiDAR, EO/IR, RADAR). This project enabled enhanced interactions via VR/AR interfaces and human-in-the-loop analytics. These efforts support dynamic decision-making and transparency under high-stakes constraints. - Analytics, data science & interdisciplinary research:
End-to-end data and analytics projects are conducted with major industrial and public stakeholders, including ArcelorMittal, Goodyear, Luxair and the Ville de Luxembourg. For example, an urban energy analytics initiative with the City of Luxembourg integrates building metadata into ML pipelines for evidence-based retrofit analysis and planning. The group also supports continual online learning, concept drift detection, and uncertainty quantification to deliver robust insights under conditions such as distributional shift. - AI benchmarking:
A full stack AI benchmarking platform provides end-to-end support for reproducible and comprehensive AI model performance assessment. Evaluation extends beyond accuracy and latency to include, for example, robustness, counterfactual sensitivity, adversarial resilience or out-of-distribution generalization. Implemented with full experiment tracking (MLflow / Kubeflow), model fingerprinting and data lineage, this platform forms part of the EDF STORE project and is deployed with zero-trust security principles. The benchmarking process follows four stages: definition, fingerprinting, orchestration and execution. This ensures deep reproducibility and auditability. The platform also facilitates the implementation of non-standard or less-standard constructs, looking at contrastive / self-supervised pretraining, domain-specific augmentation and multi-layer evaluation (micro/component/scenario level), combined with explainability methods such as integrated gradients, saliency maps and counterfactual explanations. - Multi-Agent Systems & RAG-Compliance for Regulation Parsing:
Multi-agent LLM platforms are used for parsing overlapping financial and sustainability regulations (e.g., SFDR, MiFID II, CSRD, EU Taxonomy), identifying conflicts, synthesizing compliance summaries and enabling automated reporting. The SERAFIN CORE project (Q4 2025), for instance, explores auto-calibrated chain-of-thought reasoning and uncertainty-aware logic within complex regulatory setups. In fund management compliance, the group explores the use of retrieval-augmented LLM workflows for second-level NAV error detection and breach reporting under AIFMD Article 21 and UCITS Article 22. These include knowledge-graph-enhanced retrieval, formal verification of factuality and reliability and end-to-end orchestration of LLM-based compliance processes. - Explainability and cross-domain concepts:
In projects like the EAI FNR CORE on XAI Epistemology (Q4 2024), the group investigates when opaque AI models can be deemed epistemically authoritative, in collaboration with philosophers and AI researchers. The project delivers formal models of explanation, authority and testimony, feeding directly into regulatory and audit-aligned XAI pipelines. Mechanistic interpretability is also pursued using causal abstraction frameworks, attention-flow analysis and counterfactual mutation testing to align model logic with human reasoning and support transparent decision pipelines across domains.
Ultimately, the group aims to deliver the following expected outcomes:
- Off-the-Shelf Trustworthy AI: Modular, low- / no-code building blocks incorporating explainability, robustness, experiment tracking, automated documentation, and continuous risk assessment, while observing a strong alignment with the EU AI Act and standards for high-risk systems.
- Benchmarking-as-a-Service for Regulated Environments: Tailored benchmarking suites with adversarial protocols, live dashboards, and compositional verification for safety, performance, and compliance lifecycle monitoring.
- Framework for AI-Driven Automated Monitoring of Complex Regulatory Frameworks Monitoring: Automated pipelines for regulatory risk prediction, immutable audit trails, and compliance tracking across jurisdictions,enabling streamlined reporting and proactive breach prevention.
Scope of expertise
- Cross-disciplinary AI:
Examples that span from time series processing to RAG workflows and Fin- / Reg-Tech solutions, allowing the parsing and alignment of overlapping mandates across SFDR, MiFID II, AIFMD or UCITS. Work also extends to mobile-optimized networks for astrophotography, multimodal reasoning and retrieval-aware prompting strategies. - Business analytics & ML / DataOps:
Full-lifecycle ML development, including data imputation, feature engineering, model deployment, drift monitoring and CI/CT pipelines. Efforts extend to quantum-inspired models and other future-ready compute. - Trustworthy AI & XAI:
Development of explanation techniques at local and global levels (e.g., saliency mapping, counterfactuals, surrogate modeling), evaluated through fidelity metrics and human-in-the-loop validation to meet regulatory audit criteria. AI Benchmarking includes adversarial testing automated mitigation, resilience and deep reproducibility. The group specifically focuses on advanced metric - e.g. context-aware, scene complexity, etc., X/DRL and Hybrid AI - targeting uncertainty-aware solutions, distributional shifts, or hyper-representation, contrastive/self-supervised learning to build transferable, few-shot-ready representations. - Healthcare & legal interface:
Application of AI and sata science-oriented techniques for, e.g., radiation dosimetry in healthcare, cancer treatment setups; privacy-preserving, explainable AI for diagnostics and analytics, leveraging interpretable medical and contractual decision-making. - Collaborative AI systems:
Multi-agent architectures with consensus protocols, human-in-the-loop feedback and calibration mechanisms that maintain trust and alignment across distributed environments.
Core competences of the group lies in:
- Applied AI, mathematical and/or formal modeling, Quantum-inspired solutions: Expertise in probabilistic modeling, time series analysis, Kalman/particle filters and quantum-inspired computing for certified robustness.
- Regulatory expertise: Deep domain understanding of AIFMD, UCITS, SFDR, MiFID II, CSRD; NLP pipelines for regulatory monitoring and synthetic data generation for testing and validation.
- Explainability in regulated domains: Integrated gradients, concept activation vectors, XAI interfacing and cognitive alignment validation to support human interpretability.
- Deep Learning & time series: Incremental learning architectures with concept drift detection, model adaptation and deployment in constrained environments.
- Benchmarking infrastructure: 360 experience with Kubernetes / Kubeflow / MLflow-based MLOps stack for model analysis, tracking and deployment across heterogeneous hardware setups; extensive experience in Industry 4.0, HealthTech, and Defence domains.
By synthesizing technical depth with applied impact, the Trustworthy AI Group delivers robust, explainable and compliant AI systems that address pressing industrial, societal and regulatory challenges, positioning Luxembourg as a European leader towards trustworthy AI innovation.
