Institute of Information Technology Bogra (IITB)

This paper proposes a fuzzy inexact Levenberg–Marquardt optimization (FILMO) algorithm with a descent direction to handle the nonlinear systems influenced by the uncertain parameters. The main feature of this proposed inexact algorithm is to use the Armijo-type step size search approach via an uncertain environment. The level of inexactness search direction is controlled through the descent direction of the merit function. We establish the convergence analysis of the FILMO algorithm under the assumption of local error bound. Then, the global convergence of the FILMO algorithm is described. The FILMO algorithm is constructed using fuzzy parameters with an Armijo-type step size approach. Numerical examples are illustrated to investigate the effectiveness and efficiency of the algorithm. Then, the comparison is done with a previously existing conjugate gradient modified Fletcher–Reeves method and fuzzy inner outer direct search (FIODS) method. Furthermore, to quantify the uncertainties and sensitivity of the system, fuzzy and fully fuzzy systems are investigated through a case study.

The technology of Deepfake has rapidly evolved in today’s world, it poses significant challenges to society and individuals as it enables high realistic fake images, audios and videos. There is increase of risks of deception, misinformation, and reputational damage due to these advancements. To counteract this emerging threat, we have explored Vision Transformer (ViT)-based models for deepfake detection, leveraging deep learning techniques. Our study implements ViT models —ViT-B-16 trained on datasets of 5,000 images. A Flutter-based application is developed to classify uploaded images as real or fake, providing a prediction confidence score. Experimental results indicate that the ViT-based models achieve promising detection performance, with the highest accuracy reaching 87.33%. Our research highlights the importance advanced architectures in improving deepfake detection techniques. The study of Vision Transformers, showcase the potential in tackling deepfake challenges. Our research contributes to the ongoing and future efforts to enhance the deepfake detection techniques and mitigate its social, personal and environmental impacts

The transportation sector is a significant contributor to global emissions, with road transport as a major contributor. Traditional emission control strategies such as carbon taxes, cap-and-trade systems and policy regulations often lack an effective mechanism for tracking individual vehicle emissions in real time. This paper proposes a SUMO Simulation based Carbon Credit Allocation System that collects real time data from simulation to monitor and quantify emissions per trip. The system creates a virtual representation of vehicles that allows precise emission tracking. It is facilitated with a credit-based incentive mechanism that rewards eco-friendly driving behaviors. The project ensures accuracy in emission estimation, while the credit allocation encourages sustainable transportation choices. Additionally, the system enhances transparency by providing data-driven insights for both policymakers and individuals via dashboard. Testing on two major factors demonstrated that high-speed driving results in 44% more CO₂ emissions compared to low-speed driving, whereas rough driving styles generate significantly higher emissions than normal driving. By leveraging real-time processing and adaptive learning models, the proposed system operates with high efficiency. It ensures scalable and accurate emission assessments, making it a capable solution for sustainable green mobility

With the prevailing digital era, students are distracted by social media, resulting in poor academic concentration. Our app solves this problem by turning learning into a recreational activity that is fun and accessible. Harnessing the strength of cutting-edge AI technology, the platform boasts an AI-driven chatbot that offers instant doubt-resolution, providing learners with a speedy and accurate academic answer. Besides, a social area allows users to collaborate, exchange information, and communicate through posts, images, and videos, creating an active and engaging learning community. Unlike conventional social media, our app balances learning with socialization, keeping learners motivated and inspired towards their academic pursuits. By combining AI powered support with a collaborative community, we seek to establish an effective digital platform that promotes learning, inspires peer-to-peer collaboration, and creates a strong learning community

Multimodal generative AI is rapidly transforming how preliminary diagnosis, triage, and patient guidance can be delivered through intelligent assistants. Unlike traditional symptom checkers or text-only chatbots, contemporary systems blend large language models (LLMs), vision-language models (VLMs), speech recognition/synthesis, retrieval-augmented gener- ation (RAG), and clinical decision support (CDS) logic to provide context-grounded, explainable, and more accessible care experi- ences. This survey synthesizes prior work on multimodal medical chatbots and digital twin concepts, reviews enabling architectures (RAG, knowledge graphs, explainable AI, and privacy-preserving learning), and positions a practical implementation that uses the GROQ API and a built-in multimodal model to accept voice and image inputs and return a voice response and a structured prescription with do’s/don’ts. Building on lessons from digital twin frameworks and multimodal diagnosis assistants, the survey proposes advanced enhancements: longitudinal patient- twin modeling, rationale-guided retrieval, clinical protocol valida- tion, medication-safety checks, bias and uncertainty calibration, and privacy-preserving synthetic data pipelines. The result is a blueprint for a safe, extensible, and clinically aligned medical AI assistant suitable for academic research and eventual clinical piloting