Authors - Hasan Ahmed, Ram Singh Abstract - The growth of digital media platforms has resulted in more disseminated falsehoods which now include elaborate AI-generated syn thetic text instead of manually created false information. The develop ments create major obstacles which disrupt both information trustwor thiness and public confidence. The research presents a High-Accuracy Misinformation Detection Hybrid Transformer Framework which uses BERT and RoBERTa models within an ensemble learning system. The system undergoes initial training on WELFake dataset which serves as a standard benchmark collection that contains equal proportions of au thentic and fraudulent news articles derived from both verified and un verified sources. The framework achieves adaptability through its in cremental updating process which incorporates contemporary headlines and machine-generated content. The weighted fusion mechanism merges probability results from both transformer models to decrease model spe cific bias while strengthening the system’s classification ability. The sys tem shows better results than single transformer setups and operates through a web-based system which provides immediate misinformation assessment. The study results show that using ensemble modeling to gether with scheduled model updates creates an efficient method for tackling the ongoing emergence of synthetic misinformation.
Authors - Gagani Kulathilaka, Inuka Gajanayake, Guhanathan Poravi, Saadh Jawwadh Abstract - In modern digital environments, organizations require intelligent sys tems to manage complex workflows and decision-making. Unlike most of the task management systems that are manual and give no feedback and even lack competence; this leads to poor prioritization, deadline been missed and poor com munication between teams. Thus, IntelliTask is an intelligent system of dealing with tasks, which is AI-powered and, consequently, is context-aware, giving it an edge to enhance the quality of the working processes of the people using the system (both individuals and businesses), enhancing the prioritization, and im proving the productivity. The IntelliTask platform is machine-learning models, predictive analytics, and dynamic scheduling based on identifying key tasks to balance the workloads and the cognitive load on users without the user having to engage in the task. The solution will enhance the rate at which the tasks are ac complished, making informed decisions and will bring flexibility on what task management systems will be established in the future in enterprises.
Authors - Umar Ali R, Payas Khan H, Nouriensha N, Nithish Kumar S, Nisha M Abstract - An effort to calculate the infinite value of circumference ratio is made in this paper. Instead of being made of countless infinitesimals, a given circle is parts of an new defined infinity that is single magnitude continuum derived from the change in direction that indicates that there is a jumping from finiteness to infinity .This single magnitude continuum is the accumulations of infinitely many finite magnitudes and can never be achieved by forever extending continuously finite magnitudes.The change in direction implies that infinite length (i.e. infinite distance) can be defined as two parallel lines that never intersect ,which denotes that only the terminal end of the first straight line is meaningful when extending towards infinite distance, and this terminal end is defined as infinite length, which is a magnitude that cannot be discussed any magnitudes outside of it. When the first straight line extends to infinite distance, its one-dimensional feature will be lost and become an infinite dimensional magnitude, which is determined by the change in direction.The infinite value of circumference ratio is this new defined infinity.
Authors - Sarah Rahim, Guhanathan Poravi Abstract - Compressed-domain audio steganography poses a critical foren sic challenge in modern VoIP systems, particularly within low-bitrate codecs. Traditional deep learning models often lack interpretability and struggle with low embedding rates. This paper introduces AUSPEX, a lightweight forensic framework ( 170k parameters) optimized for uni versal compressed audio steganalysis. A novel three-channel tensoriza tion strategy is proposed; incorporating raw bits, temporal derivatives, and bit stability to amplify subtle embedding perturbations. A non trainable high-pass residual stream further enhances sensitivity to first and second-order temporal noise. To ensure forensic transparency, a dual level explainability framework integrates intrinsic spatial attention with post-hoc Integrated Gradients, providing bit-level evidence attribution. Experiments demonstrate detection across CNV and PMS algorithms at low embedding rates. AUSPEX advances the field by unifying efficient, edge-deployable detection with rigorous human-centric forensic interpretability.
Authors - Md. Mehedi Rahman Rana, Md. Anisur Rahman, Kamrul Hasan Talukder, Syed Md. Galib Abstract - The adoption of AI in the law sphere on a larger scale has left new opportunities of case analysis and verdict prediction as well as legal texts interpretation with the help of the robot. However, the existing Legal Judgment Prediction (LJP) systems are submissible to implicit data bias, which contains adult information on such delicate aspects as gender, caste, occupation, and socio-economic status. These biases may result in ethically unsound and unreliable forecasting, which is a vital issue in high stakes judicial settings. This work provides a Bias-Aware Legal Case Classification and Judgment Interpretation architecture that enables improved levels of fairness, interpretability and contextual reliability in legal decision support systems. The bias-sensitive preprocessing pipeline proposed combines the Named Entity Recognition and zero-shot and legal-specific bias-tagging. These two types of vocabularies are used with a dual-encoder framework based on LegalBERT on bias-masked data and BERT on unmasked data in order to trade-off legal reasoning with controlled demographic awareness. Representations in a gating-based fusion mechanism are combined in advance to make final classification. The system is set to work on the real case documents of the Indian laws based on the publicly available repositories. Instead of substituting the jurisdictional powers, the framework is intended to deliver ethical, transparent, and contextually sensitive support to the legal practitioners. The research is relevant in the history of responsible AI, as it focuses on the issues of fairness and interpretability in the field of automated legal analytics.
Authors - Leonardo Juan Ramirez Lopez, Cristian Santiago Cruz Jimenez, Johan Sebastian Ayala Gaitan Abstract - Ongoing technological progress has significantly increased global energy demand, particularly in rapidly developing economies, a trend further intensified by continuous population growth. Although improving energy efficiency is a universal objective, it remains an unresolved challenge. Advances in science and engineering have enabled the creation of diverse energy-harvesting technologies that utilize established non-conventional sources— such as solar, wind, thermal, hydro, piezoelectric, electromagnetic, and bio-battery systems—as well as emerging concepts like rectenna-based collection. This study aims to present a comprehensive evaluation and comparison of these technologies by examining their energy sources, availability, conversion principles, infrastructure needs, production costs, performance outputs, application domains, overall efficiency, harvesting capacity, constraints, resource characteristics, and commercial feasibility. By offering a systematic comparison, the authors seek to clarify the strengths of each approach while also highlighting the practical challenges involved in applying them to meet present and future global energy demands through both existing and prospective alternative energy solutions. The main objective of this paper is to systematically evaluate and compare a wide range of energy harvesting technologies—spanning established non-conventional sources and emerging concepts—by analyzing their operating principles, resource availability, infrastructure requirements, cost, efficiency, performance, limitations, and practical applicability, with the aim of identifying their strengths, challenges, and potential contributions toward meeting current and future global energy demands through sustainable alternative solutions.
Authors - Asmit U. Patil, Sneha Jadhav Mane, Swati Suryawanshi, Prerana Mahajan, Priya Sharma, Smita Shedbale, Dhanaraj S. Jadhav, Supriya Mane Abstract - Inference latency remains a critical bottleneck in deploying large language models, for real-time and resource-constrained environments. Prior work has proposed latency formulations that express latency as a function of key parameters. However, they often assume a linear dependence on sequence length, which fails to generalize to tasks involving significantly longer sequences, such as document-level language modeling, long-context retrieval, or time-series forecasting, where latency scales nonlinearly and unpredictably. This paper addresses the limitations of existing latency formulations by proposing three complementary enhancements to improve generalization across varying sequence lengths. First, we introduce a nonlinear term for sequence length, capturing the superlinear growth in latency observed in transformer-based architectures due to quadratic attention mechanisms and memory overhead. Second, we propose a sequence-length-dependent scaling factor for the sequence length parameter itself, allowing the model to adaptively adjust its sensitivity based on empirical latency profiles across different tasks and hardware configurations. Third, we incorporate an empirical correction term enabling calibration of the latency model to account for hardware-specific and implementation-level nuances. By explicitly modeling the nonlinear and context-sensitive behavior of sequence length, our approach offers a more faithful representation of latency dynamics. This work lays the foundation for more adaptive and hardware-aware latency estimation frameworks, with implications for model deployment, scheduling, and cost optimization in production systems. We conclude by discussing future directions for integrating dynamic profiling and reinforcement learning to further refine latency predictions in evolving runtime environments.
Authors - Felipe M. Coelho, Margarida N. P. dos Santos, Jeziel M. Pessoa, William A. P. de Melo, Joel C. do Nascimento, Carlos A. O. de Freitas , Debora R. Raimundo, Vandermi J. da Silva Abstract - The transition from 4G to 5G networks, particularly in Non Standalone (NSA) deployments, introduces new challenges for the energy effi ciency of mobile devices, as they must maintain simultaneous connectivity with LTE for signaling while using 5G NR for high-speed data transmission. To ad dress this issue, this work proposes a hybrid artificial intelligence approach for predicting current consumption that combines conventional deep learning with neuromorphic computing principles. Real-world telemetry data are first pro cessed using convolutional layers and bidirectional LSTM units to capture spa tial and temporal patterns, and the resulting representations are then converted through rate coding and provided to a Spiking Neural Network (SNN). The model is trained using a hybrid strategy that integrates Spike-Timing Dependent Plasticity (STDP) with surrogate gradients, together with a custom loss function designed to emphasize prediction accuracy during high-demand periods. Experimental results show that the proposed model achieves an RMSE of 0.1164 mA, representing a 6.3% improvement compared to standard Recur rent Spiking Neural Network (RSNN) approaches, indicating its ability to cap ture abrupt variations in power consumption typical of 5G NSA environments.
Authors - Udayamoorthy Venkateshkumar Abstract - This paper focus on dual axis solar panel tracking system using Brushless Direct Current motor (BLDC), in which rotor position estimation along azimuthal angle and elevation angle is predicted using incremental en coder. The physical kinematics and dynamics parameters which are non-linear in nature is converted to linear form and processed in conventional estimated kalman filter (EKF) algorithm. The physical process noise covariance value Qk and measured noise covariance value Rk is estimated from conventional EKF predicted value, using sliding window method. Smoothing factor λ is used for quick convergence and tuning factor to estimate the process noise covariance. The simulation is performed using Python and results shows rotor position es timation along azimuthal angle is improved by 50% and 55% along elevation angle. Dual axis estimation error convergence during dynamic tracking along azimuthal angle is reduced by 66% and along elevation angle is reduced by 70% when compared to conventional EKF algorithm.
Authors - Ananya Kale, Aditi Jaikar, Shravika Hamjade, Neeta Maitre, Rashmi Apte, Mangesh Bedekar Abstract - Singer identification is a challenging task because of pitch and me lodic variations, tempo, vibrato, and adaptive singing styles. This paper propos es a novel approach towards singer identification and classification by adapting a model originally meant for speaker recognition. Specifically, this work utiliz es vector representations extracted from a pretrained Speech Brain Emphasized Channel Attention, Propagation and Aggregation in Time Delay Neural Net work (ECAPA-TDNN) model. The research pipeline processes a custom curated dataset of four prominent Indian playback singers into fixed, 8 second audio clips, with mono channel sampled at 16 kHz and exported as wav files. The Speech Brain Emphasized Channel Attention, Propagation and Aggrega tion (ECAPA) encoder transforms these labelled clips into fixed embeddings which are unique vector representations of voice characteristics of each audio clips. A suite of classical machine learning classifiers is trained on these em beddings. The study evaluates four of them namely, Logistic Regression, Sup port Vector Machines, Random Forests, and a Multi-Layer Perceptron (MLP). The MLP achieved the highest accuracy of 99.38% on held-out test data. Sup porting this result, both confusion matrix analysis and t-SNE projection clearly demonstrate clear cluster separation based on individual singer identities. These findings thus collectively validate that ECAPA embeddings contain sufficient identity-bearing structure on a singing voice. This analysis thus concludes that adaptation of speaker recognition models with appropriate classifiers is a great ly effective and efficient approach for singer identification.
Authors - Mohammad Kaif, Anshika Banyal, Rohitashwa Dey, Shashi Mehrotra Abstract - A Natural Language Interface (NLI) lets users ask questions to get data from a database without having to learn a new language like Structured Query Language. Structured data with text is needed for many applications in many fields, such as search engines, customer service, and healthcare. There are many problems that have been studied, such as the popularity of relational databases, the complexity of configuration, and the processing needs of algorithms. Translating plain language into database queries is only one of these problems. The resurgence of natural language to database queries research is driven by the increasing prevalence of querying systems and speech-enabled interfaces. The last poll on this topic was done six years ago, in 2013. As far as we know, there hasn't been any recent research that looks at the best natural language translation frameworks for both structured and unstructured query languages. We examined 47 frameworks from 2008 to 2018 in this report. 35 of the 47 were very useful for what we do. There are three kinds of SQL-based frameworks: connectionist, symbolic, and statistical. There are two types of NoSQL-based frameworks: semantic matching and pattern matching. After that, these frameworks are judged based on their language support, heuristic rule sys-tem, interoperability support, dataset scope, and overall performance. The results showed that 70% of the work to make natural language queries work with databases has been done for SQL. NoSQL languages like SPAROL, CYPHER, and GREMLIN only account for 15%, 10%, and 5% of the work, respectively. It has also been found that most of the frame-works only work with English.
Authors - Avisek Sharma, Arpita Dey, Buddhadeb Sau Abstract - The increasing adoption of intelligent transportation systems has high lighted the importance of preventive vehicle safety mechanisms that address critical human factors such as unauthorized access, alcohol impairment, and driver fatigue. This review presents a structured analysis of recent research on automated vehicle access and driver alert systems that integrate biometric au thentication, alcohol sensing, and vision-based drowsiness detection. Embedded platforms, particularly Raspberry Pi– based implementations, are examined alongside computer vision techniques for facial and eye-state analysis and MQ series sensors for alcohol detection. The study reviews and compares commonly used algorithms, including classical feature-based methods and deep learning ap proaches, in terms of detection accuracy, computational requirements, and real time suitability for embedded environments. Communication strategies for alert generation and remote notification are also discussed. The review identifies key challenges related to multi-module system integration, robustness under varying illumination conditions, and long-term sensor calibration. It concludes that an integrated, low-cost, and real-time embedded framework offers a practical and scalable approach to improving vehicular security and reducing road accidents by ensuring that only authorized, sober, and alert drivers operate vehicles.
Authors - Alena Rodicheva, Svetlana S. Bodrunova, Zaeem Yasin, Ivan S. Blekanov, Nikita Tarasov Abstract - Polycystic ovary syndrome (PCOS) is a complex of symptoms that affects many women and is estimated to affect 6 to 12% of women of childbearing age. This commonality makes it hard for healthcare professionals to give an accurate diagnosis of PCOS and thereby received adequate treatment. We created a computer system that converses with users and guides their understanding of PCOS. This system uses a language model called Ollama, which is similar to the LLaMA model. We also added a vast detailed database about PCOS categorized into 12 sections. It analyzes user questions to ensure that the responses are relevant and correct. The system was trialed with positive performance. It accurately detected PCOS related queries and formulated appropriate responses. Well, the system is very smart and can go through a huge amount of data to find for each question three most relevant answers. The most common application is augmenting LLM with scraping & performing other programming operations over the LLM to give more accurate answers than just a language model. We developed a computer program that can help PCOS patients without compromising their privacy. This system even has benefits for healthcare providers as it provides information that aids them in such treatments for women with PCOS. This project is a great example of using computer programs to help humans with PCOS and other similar things.
Authors - Lakshmi Priya G G, Gokulakrishnan. V, Nithin Joel. J, Padmalakshmi Govindarajan Abstract - Potatoes are among the most widely farmed crops globally. Healthy potato plants are crucial for the large-scale production of potato-derived foods. However, a vari ety of leaf diseases can harm potato plants, with Early Blight and Late Blight being the most prevalent. In this investigation, we employed a dataset of 1500 photos comprising healthy, early, and late blight leaves. For the diagnosis of leaf diseases, we used a transfer learning-based Ensemble Modeling. We selected Effi cientNetB0, ResNet50, MobileNetv2, and VGG16 as transfer learning models, integrating logistic regression as a meta-classifier within the Ensemble Model. We have attained 99.4% accuracy in distinguishing disease-affected leaves from healthy potato leaves, which is better than most of the recent works. For the per formance measurements, we employed accuracy, precision, recall, and F1-score. To ensure the credibility of our technique, we have integrated explainable AI (Grad-CAM) for our models, which indicates which parts of the image play a vital role in our model’s performance.
Authors - Muhamad Surya Nugraha, Dedy Rahman Wijaya, Tuntun Aditara Maharta Abstract - The widespread adoption of Kubernetes for orchestrating micro services has heightened monitoring complexity if we focus on identifying per formance degradation not visible at the level of infrastructure resource utiliza tion. In this paper, we present an application-centric AIOps framework that can be leveraged to detect “high-latency, low-resource” anomalies in Kubernetes microservices. Traditional autoscaling mechanisms that only rely on resource metrics (CPU and memory) fail to provide optimum response time with the emergence of reactive applications. The model for anomaly detection is trained using machine learning classifiers such as Random Forest, LightGBM, and Lo gistic Regression. This approach leads to a weak supervision-based approach to label datasets, with respect to Service Level Objective (SLO) violations. A course registration system is proposed to validate the application of this frame work under conditions of high concurrency and parallel simulation traffic. Ex perimental results show that the established machine learning model exhibits marked performance compared with normal threshold methods, leading to im proved operational steadiness and service robustness.
Authors - Y. Rama Devi, Panigrahi Srikanth, Devansh Makam Abstract - Large language models have shown strong potential for Arabic medical text generation; however, traditional fine-tuning objectives treat all medical cases uniformly, ignoring differences in clinical severity. This limitation is particularly critical in healthcare settings, where errors in severe cases contain higher clinical risk. In this work, we propose a severity-aware weighted loss for fine-tuning Arabic language models on medical complaint–response data. The method depends on soft severity probabilities to dynamically scale token-level loss contributions during optimization, thereby prioritizing clinically critical interactions without modifying model architectures. Experiments are conducted using the MAQA dataset, which provides Arabic medical complaints and trusted human responses. Severity labels and probabilistic scores are automatically derived using a fine-tuned AraBERT-based classifier and incorporated exclusively at the loss level. The proposed approach is evaluated across ten Arabic large language models of varying architectures and parameter scales. While standard cross-entropy fine-tuning yields only modest improvements, severity-aware optimization consistently achieves larger gains. Using a balanced weighting configuration, performance improves from 54.04% to 66.14% for AraGPT2-Base, from 59.16% to 67.18% for AraGPT2-Medium, and from 57.83% to 66.86% for Qwen2.5-0.5B, with peak performance reaching 67.18%. Overall, severityaware fine-tuning delivers improvements of up to 12.10% over non-finetuned baselines, demonstrating robust and architecture-consistent gains.
Authors - Siddharth Jha, Mayur Jaiswal, Ajay Deshmukh, Kajal Joseph Abstract - The importance of agriculture for the survival of humans and the economic stability of the world continues to grow; however, at the same time, it has also come to face many severe problems due to increasing population figures, climate change, and the loss of natural resources. The traditional techniques for crop monitoring are mostly based on manual surveys and the use of vision for inspecting crops; thus, they are regarded as too labor-intensive, time-consuming, and passive in nature, thereby becoming ineffective for managing modern large-scale farming techniques. The avail-ability of the latest technological features, such as remote sensing, Internet of Things (IoT) devices, unmanned aerial vehicles (UAVs), artificial intelligence (AI) devices, and block chain technology, has transformed crop monitoring techniques into an intelligent and proactive process for farmers to monitor crops in an efficient and precise manner. Drones play an important role in crop monitoring by using high-resolution imaging devices for detecting early crop problems, such as crop stress, pest infestations, or nutrient deficiencies, whereas IoT devices are utilized for real-time monitoring of fluctuating environment parameters, such as soil, in an intelligent manner. All these innovations help towards a high and efficient agricultural system within a sustainable environment. Hence, there are still certain limitations and hindrances faced by these advanced techniques, including high initial cost, complexity, infrastructural constraints, and data privacy, limiting these techniques for small and marginal farmers. Hence, in this review paper, a detailed review of advanced crop monitoring techniques used in agriculture is discussed; further, a critical analysis of these techniques for achieving these requirements with efficiency and standards, and an understanding of these techniques for achieving a sustainable and robust ecosystem in an agricultural system is discussed.
Authors - Neha Kriti, Arti Devi, Sarthak Srivastava, Varun Dutt Abstract - Localization in Autonomous Underwater Vehicles (AUVs) continues to be a major challenge in GPS-denied environments, where inertial drift, low visibility and uncertain motion models frequently un dermine state estimation. In this paper, we present our visual-inertial odometry framework A-KIT VIO specifically designed for resilient pose tracking underwater. The system employs tightly coupled monocular camera observations with IMU data using an Extended Kalman Filter to maintain high-rate inertial propagation along with feature-based vi sual updates to avoid drift. To address the frequent covariance mismatch during non-stationary maneuvers, we added a transformer-based module to adaptively adjust IMU process noise based on the vehicle’s immediate motion context. This method of uncertainty modeling ensures filter sta bility in scenarios where standard, fixed-noise configurations typically diverge. Evaluated within a Gazebo-based underwater simulation, the framework demonstrated consistent state estimation and bounded drift over long-range trajectories, highlighting the efficacy of adaptive covari ance for reliable underwater localization.
Authors - Fatima Batool, Farzana Jabeen, Tahira Anwar Lashari, Mehvish Rashid Abstract - Autism Spectrum Disorder (ASD) is an invisible disorder that is of ten misdiagnosed in Pakistan due to unawareness and social stigma. There ex ist multiple technological digital interventions for children with autism designed to target their social, emotional or cognitive skills. However, recent studies have overlooked on the intervention integrating the psycho-social and behavioral skills of children with autism. This mixed-method study evaluates the effectiveness of a multi-modal learning framework, SHAAZ, for psycho-social and behavioral skills enhancement of children with ASD. Employing the proposed research design, the 7 week intervention was tested on autistic children with different severity level of disorder, aged 4 to 12 years. The results revealed that across observations, there is an improvement in users performance scores. The System Usability Scale (SUS)andAppQualityandImpactEvaluationbasedonMobileAppRatingScale (MARS) scores for the designed product was 89.16 and 4.27 respectively, imply ing high usability, user engagement and a positive impact on the targeted skills of the users.
Authors - Onkar Yende, Nayan Bhutada, Mohit Thakre, Sai Khadse, Mridula Korde Abstract -Reliable stock price forecasting remains challenging due to the noisy, nonlinear, and non-stationary characteristics of financial time-series data. Traditional statistical methods and deep learning models that rely solely on raw price data often struggle to capture short-term fluctuations and evolving market dynamics. To address these limitations, this study proposes a hybrid forecasting framework that integrates causal time-domain filtering, time–frequency feature extraction, and deep learning–based temporal modeling. The proposed approach employs Savitzky–Golay and Kalman filters to sup press high-frequency market noise while preserving important price trends in a causality-aware manner suitable for real-time forecasting. Localized spectral fea tures representing transient and time-varying market behavior are then extracted using the Short-Time Fourier Transform (STFT). These enhanced time-domain and frequency-domain features are combined and modeled using a Long Short Term Memory (LSTM) network, which effectively captures long-range depend encies and nonlinear temporal patterns in financial data. The framework is evaluated using standard performance metrics, including RMSE, MAPE, and R². Experimental results demonstrate that integrating causal filtering with STFT-based features significantly improves forecasting accuracy and robustness compared to baseline models, providing a reliable and practical solution for short-term and multi-step stock price prediction.
Authors - Meixin Hu, Chuanchen BI Abstract - Speech synthesis is an important tool for improving human-computer interac tion, accessibility, and other multimedia applications. Traditional Text-to-Speech (TTS) systems have issues related to robotic tone, slow inference and lack of expressiveness. This current study presented a realization of the effectiveness of the neural TTS system using Fast Speech 2 as the underlying neural TTS sys tem. The system used in the current study was a combination of Fast Speech 2 as the underlying neural system in generating high-quality utterances and HiFi-GAN as the underlying neural vocoder. The process involves reconstructing natural-sounding text utterances in terms of mel-spectrograms by Fast Speech 2 that incorporate the use of variance adaptation in terms of pitch, duration, and energy. The implementation of natural-sounding utterances in terms of mel spectrograms is done in real-time using HiFi-GAN. The implementation of the available studies provided insights into Fast Speech 2’s effectiveness in generating mel-spectrograms in real-time and faster. The use of HiFi-GAN provided insights in generating natural-sounding utterances in real-time. The effectiveness of Fast Speech 2 in generating high-quality utterances has further stretched the poten tial use of Fast Speech 2 in virtual assistant applications, audiobooks, accessible text services, further highlighting its significance in advanced human–computer interaction systems.
Authors - Cheng Cheng, Chuanchen BI Abstract - In recent years, there has been an increase in AI - generated images. This poses a major challenge in distinguishing fabricated images from real ones. This distinction is valuable for discovering misinformation and preserving digital trust. Some deep learning models, particularly large Convolutional Neu ral Networks (CNNs), have demonstrated high accuracy on benchmark datasets like CIFAKE, but their computational requirements often in clude specialised hardware like powerful Graphics Processing Units (GPUs), which ultimately limit practical deployment. This paper explores an alternative approach that focuses on efficiency and interpretability. The CIFAKE dataset is used, but a significantly lighter CNN architecture, ResNet18 is deployed which does not require high end local GPU hardware. Furthermore, the paper applied Gradient - weighted Class Activation Mapping (Grad - CAM) not just for visu alization, but also to validate that the model learns meaningful visual features that are relevant to the classification task. This work highlights a practical method to interpret AI - generated images.
Authors - Jiayan Peng, Chuanchen Bi Abstract - With the continued growth of digital education (and multiple platforms providing education/courses), students have many things to deal with in terms of finding useful content (e.g., Lecture videos; audio files; PDF's; slides, etc) and as a result, it may be difficult to efficiently scan and gather all of this information. AutoNoteX is a tool that automatically creates notes from your spoken word using speech-to-text technology (e.g. Whisper), Natural Language Processing, and various AI agents. AutoNoteX will provide accurate transcriptions, along with structured summaries that highlight key points and provide diagrams when appropriate in order to create good, clear notes for students. AutoNoteX can support collaborative and independent learning by allowing the user to merge their notes with Google Docs or download them as PDF's. AutoNoteX also includes interactive knowledge checks that have multiple levels of difficulty (easy, medium, difficult) when answering questions and also provide a means for the student to receive instant feedback on their progress. AutoNoteX was developed using React.js for the front end and Python Flask for the backend, and is cloud-enabled (scalable; accessible via many devices; and easy to integrate into a variety of subjects) giving students the tools they need to create better notes. Overall, AutoNoteX provides a new avenue for multi-modal, AI-assisted, and personalized digital note-taking, while reducing the amount of time needed to make notes and improving student comprehension by encouraging students to participate in their learning process actively.
Authors - Qixuan Geng, Chuanchen BI Abstract - Efficient nutrient management is vital in a sugarcane cultivation to sustain the crop yields. But, the conventional practices are still reactive and imprecise often leading to improper nutrient management and yield loss. To overcome this issue, the study utilizes a multimodal AI driven framework by integrating drone-based canopy imaging and in-field soil sensors to aid in real-time nutrient deficiency detection and precise recommendation of fertilizers. UAV images are analysed using a transfer learning based Convolutional Neural Network (CNN) to locate visible deficiency symptoms and determine its severity. In order to forecast impending nutrient deficiencies, significant soil parameters (NPK, moisture, pH, electrical conductivity and temperature) are monitored continuously and processed using GRU/ LSTM- based models. The data and information from sensor networks, images and environmental context are then integrated through a fusion architecture to produce a nutrient deficiency label, severity score, and confidence measure. To ensure interpretability and agronomic safety, predictions are incorporated with crop growth stage- specific nutrient gap model that convert deficiencies into dosages of fertilizers, with alerts given on high-risk conditions and optionally permissioned fertigation control. The proposed system allows proactive, data-driven nutrient management, mitigates the risk of over fertilization, and supports scalable precision agriculture.
Authors - Md. Riaz Mahmud, Kazi Asif Ahmed, Md. Rafiqul Islam, Kabya Guha Abstract - Modeling multi-scale spatial dependencies is essential in histopathology image analysis, where diagnostically relevant patterns span cellular textures and tissue-level structures. While convolutional neural networks effectively capture local features, they struggle to model long-range interactions, and transformer-based approaches address this limitation at the cost of quadratic computational complexity with respect to spatial resolution. In this work, we propose HiSS-Fuse, a linear-time hierarchical state-space fusion framework that integrates multi-scale fea ture representations using Mamba-based selective state-space modules. The proposed architecture progressively fuses local and global contex tual information across network depths while maintaining O(L) com putational complexity, where L denotes the number of spatial tokens. Experimental evaluation on the PathMNIST benchmark demonstrates that HiSS-Fuse achieves 97.0% classification accuracy with an AUC of 0.997 while maintaining strong computational efficiency. Ablation stud ies further confirm that hierarchical fusion systematically enhances rep resentation learning. Overall, HiSS-Fuse provides a scalable and compu tationally efficient alternative to quadratic attention-based architectures for multi-scale histopathology image analysis.
Authors - Cheng Cheng, Chuanchen BI Abstract - The increasing reliance on Information and Communication Technology (ICT)-driven intelligent systems has transformed organizational decision-making processes, enabling more efficient, data-driven, and adaptive strategies. These systems, which encompass artificial intelligence, machine learning, and decision support tools, have revolutionized how businesses process and analyze vast amounts of data to inform strategic decisions (Cheng et al., 2017; Yoo & Lee, 2020). This paper presents a strategic framework for integrating ICT-driven intelligent systems into organizational decision-making, addressing key challenges such as technological compatibility, organizational resistance, and alignment with strategic goals (Patel & Sharma, 2019; López et al., 2019). The main objective of this study is to develop a comprehensive and practical framework that organizations can adopt for successfully integrating intelligent systems into their decision-making processes. The research aims to bridge the gap between existing theoretical models and practical applications by proposing a step-by-step process that involves assessing organizational readiness, selecting appropriate systems, ensuring seamless integration, and fostering continuous improvement (Ahmad et al., 2021; Pereira et al., 2021). The methodology employed includes qualitative case studies from diverse industries, supplemented with a review of relevant literature and theoretical models such as the Technology-Organization-Environment (TOE) framework (Tor-natzky & Fleischer, 1990) and the Resource-Based View (Barney, 1991). The findings suggest that successful ICT integration is contingent upon a well-planned, strategic approach that aligns technological capabilities with organizational goals and promotes an adaptive organizational culture (Brinkman & Möller, 2018). The implications of this study are far-reaching, offering valuable insights for managers and policymakers to overcome integration barriers and optimize decision-making using intelligent systems (Hossain & Kaur, 2021). This research contributes to the growing body of knowledge on ICT integration in decision-making, offering both theoretical advancements and practical guidelines for successful implementation.
Authors - Tajamul Islam, Ruby Chanda Abstract - The present study explores the online privacy concerns of young Indian consumers. Using the segmentation approach popularized by Dr Alan Wes-tin in the U.S., this study identifies the segments within Indian youth. This study is based on a survey conducted on a sample of Indian university students. Hierarchical and non-hierarchical cluster analysis techniques were applied to identify segments within young Indian consumers based on their privacy concerns. The study identified three consumer segments: highly concerned, moderately concerned, and less concerned based on online privacy concerns. The findings also reveal important differences among the three segments in terms of out-come variables such as perceived effectiveness of legal/regulatory policy, fabricating personal information, and software usage for protection. The results indicate an overall increased level of concern for online privacy among young Indian consumers. The results suggest similarities and dissimilarities with Westin’s approach. While previous research on online privacy has been chiefly based on the Western context, this study offers a window to look at the Eastern context by examining the privacy concerns of young Indian consumers, who have not been studied, and hence provides an important contribution to the existing literature.
Authors - Meixin Hu, Chuanchen BI Abstract - Secret-sharing schemes are fundamental cryptographic primitives en- abling secure distribution of sensitive information among multiple parties. Orig- inally introduced to protect cryptographic keys, they have evolved into power- ful tools underpinning modern secure multiparty computation, distributed stor- age, blockchain systems, and privacy-preserving machine learning. This review presents a systematic overview of threshold secret-sharing schemes, ramp con- structions, and secret-sharing schemes for arbitrary access structures. We discuss information-theoretic foundations, lower bounds, structural generalizations, and recent advances. Furthermore, we highlight emerging applications in distributed computing, post-quantum cryptography, and secure AI systems.
Authors - Ying Tang, Chuanchen BI Abstract - This article presents a comprehensive analysis of methods and recent research in the sentiment analysis of Uzbek-language social media posts. A balanced corpus of 100,000 posts from Telegram, Instagram, Twitter, and Facebook was constructed as the object of study, in which positive, neutral, and negative classes are equally represented. The data were subjected to thorough preprocessing steps including cleaning, normalization, tokenization, removal of stop words, stemming, and lemmatization. The evaluated models include Naive Bayes, Support Vector Machines (SVM), Conditional Random Fields (CRF), Long Short-Term Memory networks (LSTM), and transformer-based architectures such as BERT and RoBERTa. The accuracy, F1-score, and runtime performance of each model were compared. Experimental results indicate that transformer-based models achieved the highest accuracy (~92%), followed by LSTM (~90%) and SVM (~88%). Despite being a simple method, Naive Bayes served as a baseline (~78% accuracy). The literature review highlights prior research conducted in Uzbek sentiment analysis, emphasizing the importance of corpus creation and accounting for language-specific features. The results indicate that transformer models provide the highest accuracy, whereas classical methods remain competitive even in low-resource settings. The article concludes with a discussion of promising directions and potential practical applications in the field of Uzbek-language sentiment analysis.
Authors - Lankalapalli Vamsi Krishna, Santanu Mandal Abstract - The rapid advancement of generative and agentic artificial intelligence (AI) is significantly transforming research in operations management and supply chain systems. Despite the substantial increase in scholarly output in recent years, the structural evolution and thematic consolidation of this interdisciplinary field remain insufficiently mapped. This study presents a bibliometric analysis of 116 Scopus-indexed articles published between 2015 and 2025 to examine publication trends, knowledge concentration, intellectual structure, and longitudinal thematic transitions. Utilizing the Bibliometrix R package, the analysis employs performance metrics, Bradford’s Law, keyword co-occurrence mapping, thematic centrality–density analysis, and temporal evolution modeling. The results indicate accelerating research growth and increasing consolidation within core engineering-oriented journals. Intellectual clustering reveals strong integration between computational modeling, reinforcement learning, and supply chain decision systems. Thematic mapping identifies computational methods and autonomous agents as central themes, while generative AI emerges as a developing yet increasingly interconnected trajectory. Longitudinal analysis reveals a clear shift from agent-based simulation frameworks toward adaptive, autonomous, and AI-integrated operational ecosystems. The findings suggest that generative and agentic AI are becoming foundational elements of next-generation operational intelligence systems. This study provides structured insights into the maturation of AI-enabled operational research and offers guidance for future interdisciplinary investigations in autonomous supply chain intelligence.
