Authors - Prabhat Kumar Gupta, Perumal T, Karthick Pannerselvam Abstract - Generation Large language models, as well as retrieval-augmented generation (RAG), are highly performing on semantic queries, but with considerable latency as they require embedding computation, a vector similarity search, and generation at inference time. Such delays make them inappropriate in time-sensitive and domain-specific retrieval activities. In this paper, the Hierarchy Latent Retrieval Model (HLRM) which is a deterministic architecture will be introduced and able to answer semantic queries in O(1) constant time. HLRM unites hierarchical semantic routing and semantic hashing so that pre-validated units of knowledge can be directly illuminated without the need to search methods or language model informing of their existence at run time. All computationally expensive processes are done offline, which means that embedding processes or vector databases are not needed to run a query. Milliseconds-response time with very high exact-match accuracy is proved under experimental assessment on an orderly institutional knowledge environment. The findings suggest that HLRM offers an alternative of fast, interpretable, and reliable systems to the generative retrieval systems in non-random settings where precision and response latency is paramount.
Authors - Khedkar Aboli Audumbar, Uday Pandit Khot, Balaji G. Hogade Abstract - Malicious or compromised internal users can act like normal users with valid login credentials and thus become difficult to detect. As a result of their similarity to normal users, traditional methods of detecting intrusions, have difficulty identifying the subtle and changing behaviors of malicious insiders. This paper introduces a comprehensive User and Entity Behavior Analytics (UEBA) framework to help detect malicious insiders. It works by analyzing activity logs generated by the enterprise. Further it performs data cleaning and feature engineering; creating behavioral profiles for each user based upon the attributes of time, environment, and behavior. These profiles are used to model normal interaction patterns and with the DBLOF algorithm, an outlier score for each profile is created. The outlier score indicates whether or not a given user’s behavior has deviated from normal. In order to make the proposed system adaptable to changing environments over time, it utilizes deep learning algorithms to detect changes in behavior and to increase the accuracy of anomalous behavior detection. The proposed system also enables the ingestion of real-time data, the evaluation of risk, and the display of alerts in a visual format. Thus, providing the scalability and operational performance required to support large-scale organizations. Overall, the proposed system represents a reliable, modular, and understandable UEBA framework. It is capable of accurately detecting malicious insider threats and representing an efficient method for proactively mitigating risks through security operations within enterprises.
Authors - Poonam Chaudhary, Rita Chhikara, Nupur Prakash Abstract - This work addresses the challenge of Isolated Sign Language Recognition (ISLR) on mobile and edge devices, where computational resources, memory, and energy budgets are severely constrained. Existing approaches based on pixel-level three-dimensional convolutional neural networks are computationally expensive and sensitive to background variations, while recurrent models such as Long Short-Term Memory networks suffer from a sequential processing bottleneck that limits parallel execution on modern hardware accelerators. To overcome these limitations, this paper proposes a hybrid Adaptive Graph Convolutional Network (A-GCN) and Transformer architecture that decouples spatial and temporal modeling of skeletal sign representations. The A-GCN employs a learnable adjacency matrix to capture dynamic and semantically meaningful spatial relationships between skeletal landmarks, while the Transformer encoder leverages parallel self-attention to model long-range temporal dependencies without recurrence. Experimental evaluation on the 250-class Google Isolated Sign Language Recognition dataset demonstrates a Top-1 accuracy of 78.90%, outperforming a Bi-LSTM baseline by 6.96%. In addition, the proposed model achieves a throughput of 400.55 frames per second with a latency of 2.50 ms on accelerator hardware and maintained real-time performance on consumer-grade devices. These results demonstrate that landmark-based, parallel architectures enable accurate, real-time, and privacy-preserving sign language recognition suitable for deployment on standard mobile devices.
Authors - Subin Simon, Prathilothamai M Abstract - Deep learning has shown significant potential in medical image classification; however, a systematic comparison of deep feature extraction strategies for multi class diabetic eye disease assessment remains limited. This study presents a comprehensive comparative analysis of seven deep learning architectures, including conventional CNN, pretrained VGG16, Vision Transformer (ViT), Conformer, hybrid CNN ViT, and attention-augmented variants incorporating Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM). All models are evaluated under a unified preprocessing and training framework to ensure fair performance comparison.The investigation focuses on analyzing how different architectural paradigms capture discriminative local and global retinal features relevant to disease classification. Extensive experiments are conducted on public fundus image datasets using standard evaluation metrics, including accuracy, precision, recall, and F1-score. Experimental results demonstrate that hybrid and attention-integrated architectures outperform standalone CNN and transformer models. In particular, the Conformer architecture achieves the best overall performance, reaching approximately 91% classification accuracy in the four class setting (Diabetic Retinopathy, Glaucoma, Cataract, and Normal), while the CNN ViT model attains approximately 89% accuracy.These findings highlight the effectiveness of combining convolutional operations with global self-attention mechanisms for robust and discriminative feature extraction in automated diabetic eye disease classification.
Authors - M.Murugesen, Priyanka P Abstract - Deep learning–based medical image models have achieved expert level performance in GPU-based research environments [1–3]. However, relia ble deployment in real clinical systems remains challenging due to constraints related to power consumption, hardware stability, and long-term operation. While prior studies have focused on improving model architectures or hardware accelerators [4,5], relatively limited attention has been devoted to systematical ly managing the transition from GPU-based development to NPU-based de ployment environments. This study formulates the GPU-to-NPU transition as an independent deployment research problem. Rather than proposing a new model architecture, we focus on preserving functional equivalence when trans ferring a validated GPU-trained medical vision model to an NPU-based infer ence environment. The proposed framework consists of reference model fixa tion, intermediate representation (IR)-based conversion [13–15], operator com patibility management, inference pipeline alignment, and output-level function al equivalence validation. The framework is evaluated through deployment of a ResNet-50–based pa thology classification model on a commercial ATOM NPU platform. Experi mental results demonstrate a 99.1% agreement rate (991/1,000 samples) be tween GPU-based and NPU-based inference outputs, confirming consistent de cision behavior despite architectural differences. These findings indicate that deployment reliability depends more on execution environment control and preprocessing alignment than on model architecture modification. By redefining deployment as a structured research problem, this work pro vides a reproducible methodology for translating research-grade medical AI models into energy-efficient NPU inference systems under practical clinical constraints.
Authors - Jayanthi J, P.Uma Maheswari, S.Uma Maheswari, Arun Kumar, Karishma V R Abstract - The rapid migration of artificial intelligence from cloud platforms to edge-based Internet of Things environments has intensified the demand for transparent and trustworthy decision-making under severe resource constraints. While edge intelligence enables low-latency and privacy-preserving analytics, the opacity of deployed models limits user trust, accountability, and regulatory acceptance. Existing explainability techniques largely assume cloud-level resources, making them unsuitable for real-time and energy-limited edge deployments. In order to close this gap, this work develops an interpretable intelligence framework that is resource-aware and adaptable, specifically designed for limited IoT systems. The suggested approach integrates interpretability directly into the decision-making process, allowing for the generation of faithful, lightweight explanations in addition to predictions while dynamically adjusting to operational context and runtime restrictions. Further balancing local responsiveness with system- level insight aggregation and secure governance is achieved through hierarchical explanation control. Transparency, efficiency, and scalability are all in line with the framework's treatment of explainability as a fundamental system capacity. The study shows that adaptive, deployment-aware explainability can greatly improve edge intelligence's operational reliability and trustworthiness. These insights establish a foundation for building accountable and interpretable AI systems in real-world IoT environments.
Authors - Rimon Kumer Roy, Jannatul Ferdous, Kazi Lutfur Nahar Mithila, Sabbir Islam, Mohammad Zahid Hassan, Sadah Anjum Shanto Abstract - Early identification of ophthalmic disease is critical to pre serve eyesight. We present DeepEye, a stacking-ensemble framework for multi-disease classification on the Eye Disease Image Dataset (EDID, Mendeley Data). After standardized preprocessing and augmentation, f ive architectures ResNet50, VGG16, DenseNet121, EfficientNet-B4, and Vision Transformer were trained and evaluated. The final ensemble in tegrates the top base models with a logistic regression meta-learner op timized via hyperparameter tuning. On a held-out test set, DeepEye achieves 91.34% accuracy and AUC of 0.9965, outperforming all con stituent models and exhibiting stable gains across cross validation folds. Model transparency is supported with Grad-CAM visualizations that lo calize disease-relevant regions, enhancing clinical interpretability. These results indicate that combining convolutional and transformer backbones within a tuned stacking framework yields a high-accuracy, explainable approach for automated eye disease detection in healthcare settings.
Authors - Nguyen Thi Hoi, Vu Thi Anh Hong, Dang Thi Anh Tho, Dang Thuy Linh, Nguyen Khanh Linh Abstract - The increasing use of renewable energy sources has made the integra tion of Flexible AC transmission system (FACTS) devices into contemporary power systems, an important area of research. The function and effectiveness of FACTS devices in enhancing power quality and preserving stability in traditional power systems and those that significantly count on renewable energy source are comprehensively examined in this study. Variability and unpredictability brought about by renewable energy sources can negatively impact the voltage profile, particularly at high penetration levels. Devices from the Flexible AC Transmis sion System, like the Thyristor-Controlled Series Capacitor (TCSC) & Static Var Compensator (SVC), provide efficient ways to improve system stability and dy namically regulate voltage. This paper investigates a coordinated control strategy of SVC and TCSC for improving voltage profiles in a transmission network with high renewable energy integration. Using an IEEE-14 bus test system, various scenarios of renewable penetration are simulated to analyze the performance of coordinated FACTS operation. The findings show that the suggested coordinated control improves overall system dependability and power transfer capabilities in addition to reducing voltage variations and reactive power imbalances. The study highlights the importance of optimal placement and coordinated tuning of FACTS devices as a cost effective solution for enabling secure and stable opera tion of renewable-rich power grids.
Authors - Josue Piedra, Nelson Piedra Abstract - Accurate crop production forecasting is essential for sustainable agricultural planning, effective resource management, and long-term food security. Conventional statistical and regression-based models often fail to capture the complex, nonlinear relationships that exist among agro-climatic variables, soil characteristics, and crop yield [1]. To address these limitations, this paper proposes an agentic artificial intelligence (AI)–based framework for crop production analysis that integrates autonomous decision-making with machine learning and deep learning techniques. The proposed framework utilizes agro-climatic and soil parameters such as temperature, humidity, soil moisture, cultivated area, and seasonal information to model crop production behaviour. Three predictive approaches— Linear Regression, Random Forest, and CNN–LSTM—are implemented and evaluated within the agentic framework using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R2) as performance metrics. Experimental results demonstrate that the Random Forest model significantly outperforms the other models, achieving an RMSE of 0.56, MAE of 0.31, and R2 value of 0.96. These findings highlight the effectiveness of agent-driven machine learning systems in accurately modelling agricultural data and supporting intelligent decision-making for crop yield optimization.
Authors - Priyanka Halder, Anupam Sinha Abstract - This study analyzes the extent to which credibility from influencers impacts consumers' buying behavior. The focus will be on how the intention to buy impacts this relationship as the problem is being analyzed in the context of social commerce on TikTok. The study is developed within the framework of Source Credibility Theory which suggests that consumers’ perception and consequent behavior are influenced by the perceived degree of the spokesperson’s Attractiveness, Trustworthiness, and Expertise. The study employs a quantitative explanatory methodology. A purposive sampling technique was used to collect data from a sample of 100 active TikTok users who follow the provided influencer. The analyzed relationships will be quantified using Structural Equation Modelling with Partial Least Squares (SEM-PLS). The research results concluded that influencer credibility increases the intention to buy, but does not increase the purchasing decision. The intention to buy completely mediates the relationship between influencer credibility and purchasing decision. This demonstrates that influencer credibility is a significant factor in the intention to buy behavior, but it is the intention that is essential in order to convert the persuasive influence into actual buying behavior. The study contributes to digital marketing communication research by extending Source Credibility Theory to the context of short-video social commerce platforms.
