Human-Centred Machine Learning: Lighter, Clearer, Safer

Project Information

ACHILLES

Grant agreement ID: 101189689

DOI

10.3030/101189689

EC signature date 5 November 2024

Start date 1 November 2024

End date 31 October 2028

Funded under

• Digital, Industry and Space

Total cost

€ 9 616 256,25

EU contribution

€ 8 226 276,25

Coordinated by

ASSOCIACAO FRAUNHOFER PORTUGAL RESEARCH
Portugal

Periodic Reporting for period 1 - ACHILLES (Human-Centred Machine Learning: Lighter, Clearer, Safer)

Reporting period: 2024-11-01 to 2026-04-30

Summary of the context and overall objectives of the project

ACHILLES addresses the need for trustworthy, sustainable and legally compliant AI through a modular framework for Responsible AI. It targets key AI vulnerabilities—efficiency, transparency, robustness, privacy and compliance—under the principles of lighter, clearer and safer AI.
The framework follows an iterative, human-centred development cycle covering ethical requirements, data quality and fairness, privacy-preserving methods, efficient and robust model development, deployment optimisation, monitoring and accountability. A central output is the ACHILLES IDE, a specification-oriented environment with Copilot and agentic capabilities to help developers integrate bias detection, data validation, privacy checks, efficiency optimisation, explainability and governance into AI workflows.
The consortium brings together technical, legal, ethical and social-science expertise across 16 partners and 10 countries. The approach is validated in healthcare, identity verification, content creation and pharmaceuticals. ACHILLES supports EU priorities by enabling responsible AI innovation, improving compliance readiness, reducing technical and environmental costs, and strengthening trust in AI systems.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

WP1 – Data-centric AI
Activities: WP1 defined interoperability requirements, advanced identity-verification datasets and evaluation protocols, and developed methods for bias detection, data auditing and data-quality validation.
Main outcomes: WP1 delivered D1.1 new annotated datasets and evaluation resources, a first bias analysis API, the Minimum Viable Dataset concept, and initial methods for fairer, auditable and data-efficient AI.

WP2 – Privacy-preserving and GDPR-aware AI
Activities: WP2 translated GDPR and privacy requirements into operational guidance and developed approaches for anonymisation, synthetic data, federated learning, secure computation and privacy-preserving biometrics.
Main outcomes: WP2 delivered D2.1 and D2.3 a GDPR-oriented use-case analysis and compliance roadmap, and privacy-preserving assets including federated learning platforms, secure data-sharing concepts, confidential execution tools and verifiable ML workflows.

WP3 – Efficient AI and optimisation
Activities: WP3 advanced efficient and scalable AI workflows, including training and inference optimisation, cost-aware fine-tuning, hyperparameter optimisation, multi-agent execution and infrastructure recommendation.
Main outcomes: WP3 produced methods for explainability-informed optimisation, sparse inference, cost-aware retraining, adaptive scaling and infrastructure optimisation, and established an initial multi-agent basis around SOP Lang, AchillesAgentLib, Agent Skills, Ploinky and MRP-VM.

WP4 – Legal, ethical and human-values framework
Activities: WP4 mapped legal and ethical requirements, developed ethics protocols, conducted legal research on trustworthy and privacy-preserving AI, and applied Value-Sensitive Design through interviews, surveys, workshops and use-case engagement.
Main outcomes: WP4 delivered D4.1 and D4.4 translated legal and ethical requirements into practical guidance, generated insights on values and ethical tensions, and created a basis for integrating compliance and human-centred design into the IDE, Copilot and validation activities.

WP5 – Explainability, robustness and monitoring
Activities: WP5 developed methods for explainable AI, uncertainty quantification, robustness assessment, continuous monitoring, hallucination detection and automated reporting for traditional and generative AI workflows.
Main outcomes: WP5 produced assets for interpretable and auditable AI, including concept-based explainability methods, robustness and uncertainty tools, face-morphing evaluation resources, LLM monitoring specifications and Cardsmith for semi-automated AI documentation, risk assessment and governance reporting.

WP6 – ACHILLES IDE integration
Activities: WP6 defined the user, technical and architectural requirements for the ACHILLES IDE and advanced components for agent deployment, development, orchestration, MCP-based interfaces, governed LLM access, observability and structured agentic execution.
Main outcomes: WP6 delivered D6.1 and prototypes of Ploinky, AchillesAgentLib, SOP Lang, Soul Gateway and MRP-VM, creating the first operational environment to expose, compose, test and validate ACHILLES outputs through APIs, plugins, Copilot workflows and use-case scenarios.

WP7 – Use cases and validation
Activities: WP7 refined the Healthcare, SCRIPTA, ID Verification and HERA use cases by updating requirements, data considerations and validation needs, and aligning them with technical, legal and ethical work.
Main outcomes: WP7 delivered D7.1 and a pilot-preparation methodology covering use-case pipelines, Trustworthy AI test cases, asset applicability assessments and pilot-planning templates, providing the basis for use-case-driven validation and D7.2.

WP8 – Dissemination, communication and exploitation
Activities: WP8 established dissemination, communication, stakeholder engagement, clustering and exploitation activities, including strategic plans, communication channels and materials, European networking, and exploitation/IPR monitoring.
Main outcomes: WP8 delivered D8.1 and D8.3 launched the website, visual identity, LinkedIn channel and communication kit, achieved initial dissemination and networking results, and created a framework to track Key Exploitable Results, IPR conditions and exploitation risks.

WP9 – Project coordination and management
Activities: WP9 established governance and management through consortium meetings, internal communication and document-management procedures, and technical, financial and risk-monitoring processes.
Main outcomes: WP9 delivered D9.1 and D9.4 providing structured procedures for coordination, quality assurance, risk management, data governance and efficient project execution.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

By the end of RP1, ACHILLES had moved from conceptual design to concrete foundations for trustworthy, privacy-preserving, efficient and human-centred AI. Results include interoperability specifications, datasets and evaluation resources, bias and data-auditing tools, GDPR and ethics guidance, privacy-preserving learning methods, efficient AI techniques, explainability and robustness methods, monitoring and reporting tools, and the first ACHILLES IDE components.
A major achievement is the initial IDE integration environment, enabling project outputs to be exposed, composed and validated through APIs, plugins, agentic workflows and Copilot support. The four use cases were refined into operational validation structures, creating the basis for realistic testing in RP2.

Potential impacts include improved AI governance, reduced compliance and documentation burden, stronger integration of legal and ethical requirements into technical workflows, more efficient use of computing resources, and increased trust in AI systems. Further uptake will require technical integration, use-case demonstrations, prototype refinement, regulatory and standards alignment, user-centred validation, and exploitation/IPR support.