D3.3 Initial AI tools for multilingual interaction and automatic generation of AI models

This deliverable presents the initial AI tools co-designed and developed within the ALFIE project, with a primary focus on the AutoML and AutoML+ engines and their role in generating unbiased, ethically sound, and EU-compliant AI models. It outlines the technical foundations, operational workflows, and integration mechanisms that enable robust, transparent, and multilingual AI-driven solutions across ALFIE’s trial scenarios. At its core, the document details the end-to-end model generation process implemented by the AutoML+ engine, explaining how user requirements, expressed in any supported language, are automatically understood, processed, and translated into optimized AI pipelines. This multilingual capability allows users to interact with the system in their preferred language, after which the AutoML engine interprets and acts on the inputs, and subsequently translates the outputs back for the user. The document further describes the distinction between AutoML and AutoML+ and elaborates on the state-of-the-art techniques embedded in the platform, including transfer learning, hyperparameter optimization, meta-learning, and multi-objective optimization. These techniques enable efficient model selection, improved performance, and adaptability across diverse use cases, while balancing accuracy, fairness, interpretability, and computational efficiency. A complementary contribution of this deliverable is the presentation of the multilingual interaction framework. The architecture supporting both user interface multilinguality and multilingual natural language interaction is described in detail, alongside the supported languages for both the chatbot and the Agentic Core. The conversational interface is designed to simplify complex AI processes, enabling users to interact with the system intuitively and transparently, regardless of linguistic background. The document further introduces ALFIE’s ethical model generation framework. It explains how the AutoML engine incorporates drift and bias detection mechanisms prior to model training, ensuring that datasets are assessed for fairness and representativeness before pipeline construction. In addition, concept drift detection and adaptive model updating mechanisms are applied post-deployment to maintain performance, compliance, and trustworthiness in dynamic environments. Transparency and explainability outputs, as well as comprehensive traceability and logging mechanisms, are embedded throughout the workflow to support accountability and alignment with EU regulatory requirements. The application of AutoML within the ALFIE trial contexts, website accessibility, autonomous vehicle bias detection, and compliance screening, is also discussed, demonstrating how the platform’s capabilities are operationalized in real-world scenarios. These examples illustrate how automated, ethically guided model generation supports diverse domains with varying technical and regulatory constraints. Finally, the deliverable describes how the AutoML engine works with other ALFIE components, including the AutoDW, the Agentic Core, ETD-Hub, XAI and semantic knowledge graph. These components provide contextual, semantic, and meta-level information that informs and enhances model generation, ensuring coherence across the platform. Through this architecture, ALFIE enables users to create robust, transparent, and unbiased AI models in a structured, efficient, and ethically grounded manner.