The increasing complexity of public procurement regulations poses significant challenges for public administrations in accessing, interpreting, and managing regulatory information efficiently. In Morocco, this challenge is amplified by the large volume of legal documents, frequent regulatory updates, and the presence of numerous scanned archives. To address these issues, this paper proposes a sovereign Artificial Intelligence framework based on a Retrieval-Augmented Generation (RAG) architecture for regulatory knowledge management within public administration. The proposed solution operates entirely in an On-Premise environment, ensuring data confidentiality and digital sovereignty. It integrates an Optical Character Recognition (OCR) pipeline for processing scanned documents, a hybrid retrieval mechanism combining semantic and lexical search, and a locally deployed Large Language Model (Llama-3-8B-Instruct) for context-aware answer generation. The framework was evaluated using a corpus of more than 1,200 regulatory documents and a Golden Dataset composed of 50 expert-validated question-answer pairs. Experimental results achieved a Recall@5 of 92.4%, an MRR@10 of 0.94, and a semantic similarity score of 92.4%, outperforming conventional retrieval approaches while maintaining low response latency. The results demonstrate the potential of sovereign generative AI to enhance regulatory knowledge management, improve information accessibility, and support decision-making processes within Moroccan public administration.
Nihal Sajjaa, Hayat Bihri, Nasreddine Haqiq et al.· EPJ Web of Conferences· 0 citations
Automated classification of cardiac pathologies from cine-MRI remains a clinically significant challenge due to inter-patient morphological variability. This study presents a comparative evaluation of deep learning architectures and proposes an explainable ensemble framework for cardiac disease classification using the publicly available ACDC dataset. Seven pre-trained models were fine-tuned via transfer learning VGG16, MobileNet, EfficientNet, GoogLeNet, ResNet18, DenseNet, and Vision Transformer (ViT) on 1,468 cine-MRI images (80/20 train-test split). A soft voting ensemble combining the top-performing architectures was developed to improve generalization and diagnostic robustness. Among individual models, VGG16 achieved the strongest performance (accuracy: 97.28%, F1-score: 0.9643, precision: 0.9561, recall: 0.9726, specificity: 0.9726). The proposed ensemble model consistently outperformed all standalone architectures, yielding 98.23% accuracy, F1-score of 0.9762, precision of 0.9753, recall of 0.9771, and specificity of 0.9771, with a clinically relevant reduction in false negatives and an AUC of 0.9971. Explainability was ensured through complementary post-hoc analyses using SHAP, LIME, and Grad-CAM, collectively confirming anatomically coherent and clinically meaningful decision patterns. These results demonstrate that architectural diversity combined with probabilistic aggregation constitutes an effective and interpretable strategy for reliable cardiac MRI diagnosis in clinical decision support systems.
Soukaina Ait Ouaoures, Hayat Bihri, Salma Azzouzi et al.· EPJ Web of Conferences· 0 citations