University of N'Djamena

Large language models excel at generating plausible responses but often produce factually incorrect answers in high-stakes financial analysis, leading to regulatory violations and financial losses, a critical challenge for deploying AI systems in production. Traditional Retrieval-Augmented Generation (RAG) systems rely on deterministic embeddings that cannot quantify retrieval uncertainty, resulting in overconfident but unreliable answers for complex financial queries. We introduce Bayesian RAG, a principled probabilistic framework that integrates epistemic uncertainty quantification directly into retrieval using Monte Carlo Dropout, bridging the gap between theoretical rigor and practical deployment. Our approach computes distributional embeddings for queries and documents, enabling a Bayesian scoring function Si = μi−λ·σi that balances semantic relevance against uncertainty. Comprehensive evaluation on Apple and Microsoft 2023 10-K reports demonstrates substantial improvements: 93.1% accuracy with significant gains in Precision@3 (+20.6%), MRR (+22.7%), and NDCG@10 (+25.4%) over BM25 baselines, plus 26.8% better uncertainty calibration. Critically, Bayesian RAG successfully extracts precise financial figures ($211.915B Microsoft, $383.285B Apple revenue) where traditional methods fail, reducing hallucination by 27.8%. Bayesian RAG advances uncertainty quantification in retrieval systems through principled Monte Carlo Dropout integration, establishing theoretical foundations for uncertainty-aware information retrieval. The modular design enables seamless integration with existing RAG pipelines, making it immediately deployable in production systems for risk-aware AI applications in finance, healthcare, and regulatory compliance.

This study addresses the challenges of diagnosis and prognosis in complex dynamic systems, focusing on an industrial MIG/MAG robotic welding application. A robust framework based on Recurrent Neural Networks (RNNs), specifically LSTM and GRU architectures, was developed to analyze multivariate time-series data from real-world sensor measurements for fault detection and remaining useful life (RUL) prediction. The approach incorporates a post-hoc attention mechanism to enhance interpretability by identifying the most influential variables and time windows contributing to each prediction. A sliding-window method (50 historical steps and a 10-step prediction horizon) was employed, and the models were evaluated against various benchmarks, including decision trees, feedforward neural networks, random forests, 1D CNNs, CNN–LSTM, and transformer architectures. Experimental results demonstrate that GRU and LSTM models outperform all baselines in both diagnostic and prognostic tasks, achieving up to 94% diagnostic accuracy and inference times of less than 100 ms, making them suitable for real-time deployment. The attention mechanism consistently identified key degradation signatures in more than 85% of fault sequences, offering valuable insights for domain experts. The proposed RNN-based framework therefore combines predictive accuracy, real-time performance, and interpretability, providing a scalable and effective solution for intelligent maintenance in industrial robotic systems.

This study uses technological infrastructures to inform evidence-based policy design, addressing the conundrum of agricultural productivity and agro-environmental sustainability in Africa. The investigations are done using the Instrumental Variable Two-Stage Least Squares (IV-2SLS) strategy to control for potential endogeneity, covering the period 2000-2020. The findings show that farmers, in their pursuit of greater productivity, often adopt unsustainable agricultural practices, which degrade agro-environmental quality through emissions of nitrous oxide and methane gases. The findings remain consistent after considering the specific cases of crop production and animal agriculture. Similarly, results indicate that technology-infrastructure reduces nitrous oxide and methane gas emissions. Moreover, the negative marginal effect indicates that the indirect benefits for sustainable agriculture provided by integrating technological infrastructure outweigh the adverse impacts on agricultural sustainability. These findings suggest that policymakers should promote the integration of technology infrastructures into the agricultural sector, as they serve as effective tools for enhancing agro-environmental sustainability.

This study examines the linear and the non-linear effects of international trade taxes on access to clean cooking technologies in Sub-Saharan Africa between 2000 and 2023. The study adopts several robust estimation strategies, including the Driscoll-Kraay cross-sectional dependence estimation, the two-step system GMM approach, and the Dynamic Panel Thresholds strategy. The findings reveal a dual effect: while international trade taxes exert a positive linear impact at moderate levels, they produce a negative non-linear effect once the tax rate surpasses a dynamic threshold of 5.039. The estimated dynamic threshold of 5.039% marks the turning point beyond which trade taxes negatively affect clean cooking access. The results are consistent across rural and urban populations, low and middle-income economies, as well as between countries with low versus high energy import dependence, with the non-linear effect most severe among high importers. Complementary factors such as internet penetration, trade openness, governance quality, and domestic credit significantly enhance access to clean cooking fuels and technologies. These results underscore the need for "smart" taxation policies that balance fiscal space generation with household affordability. Policy recommendations emphasise differentiated tariff regimes, strengthened credit markets, improved governance, and regional trade cooperation to accelerate universal clean cooking access in SSA.

Reducing air pollution in the transport sector is becoming a necessity for multiple environmental problems. This article highlights an optimal vector control strategy for an induction motor intended for electric vehicles., as a sustainable and less polluting alternative, aiming to solve the problem of adaptability to the varying dynamics of vehicle operation. The scientific contribution of this article proposes a model of the induction motor associated with an inverter controlled by pulse width modulation (PWM), then an optimal vector control with Proportional Integral (PI) and Artificial Neural Network (ANN) controllers. Four simulation scenarios are developed to better highlight the impact of the proposed strategy. The results obtained show a significant improvement in the adaptability of the system with a rise time of 0.08s, a response time of 0.97s and an overshoot of 9.1% for the PI controller. Also, show a rise time of 0.043s, a response time of 0.17s and an overshoot of 2.1% for the Artificial Neural Network controller. Its results presented in this work, confirm that this approach does not just improve the simple control of the motor, but opens the way to a fine and adaptive control of the dynamic behaviors of the system, thanks to the optimal vector control.