EXCALIBUR
EXplainable and demoCrAtized pipeLInes for unBiased, trUstworthy, and Responsible human-centric AI
About the Project
- • EXCALIBUR develops a human-centric framework for Trustworthy and Responsible AI, using LLMs to generate clear, model-agnostic explanations and fairness interpretations for AI systems.
- • It bridges regulatory principles and technical requirements, translating transparency, fairness, and accountability into actionable system specifications and ethical guidelines.
- • The project delivers an open-science platform and toolset with human-friendly visualizations, enabling understanding, interaction, and feedback, even for non-experts.
- • Validation combines in-lab experimentation (method development and fine-tuning) with in-the-wild case studies with non-expert users, following a Citizen Science approach.
- • Initial use cases focus on the wearables domain, where trust, clarity, and fairness are especially critical.
At a Glance
- Project Title
- EXCALIBUR: EXplainable and demoCrΑtized pipeLInes for unBiased, trUstworthy, and Responsible human-centric AI
- Duration
- 36 months
- Total Budget
- €300,000
- Host Institution
- Aristotle University of Thessaloniki (AUTh)
- Principal Investigator
- Athena Vakali
- Scientific Area
- SA5 "Mathematics & Information Sciences"
- Scientific Field/Subfield
- 5.2 "Computer and information sciences" - 5.2.7 "Artificial intelligence, intelligent systems, multi-agent systems"
Core Components
EXCALIBUR is structured around three core components in response to the growing need for trustworthy and responsible artificial intelligence.
Framework
A research-driven framework that leverages Large Language Models (LLMs) as advanced, model-agnostic explainers and fairness evaluators, integrating insights from established explainability and fairness approaches into clear, human-understandable outputs.
Toolset
An open-source toolset and visualization platform that embeds the framework into real AI pipelines and presents explanations and fairness reports in a transparent, user-friendly way, enabling human insight, interaction, and feedback.
Evaluation & Case Studies
A two-track validation approach combining in-lab experimentation for method development and fine-tuning with in-the-wild case studies involving stakeholders, assessing usability, trust, clarity, and overall impact through iterative evaluation.
Key Pillars
The foundational principles that guide EXCALIBUR's approach to trustworthy AI.
Regulatory-to-Technical Mapping
The systematic translation of Trustworthy AI principles and regulatory requirements into concrete technical specifications and ethical guidelines that shape the framework and platform design.
Model-Agnostic Explainability
LLM-based components that produce human-like explanations and contextual fairness interpretations across tasks and metrics, supporting the identification and understanding of potential sources of bias in data and models.
Open Tools & Stakeholder Validation
An open infrastructure for reuse (covering code, tools, and platform components) combined with continuous evaluation involving researchers and non-expert users, ensuring practical usefulness and societal relevance.
Get Involved
EXCALIBUR follows a citizen-science-inspired approach. Stay connected for upcoming opportunities to participate in our in-the-wild case studies and help shape the future of trustworthy AI.