Computer Engineering, Computer Science, Information Systems, Computer Science Applications
15
Scopus Publications
Scopus Publications
Adaptive Ensemble Meta-Learning for Enhanced Student Focus Prediction Umesh R, Suriya T S, Vishal S, Ayyanar Muthu T Iccids 2026 9th International Conference on Computational Intelligence in Data Science, 2026 Accurately predicting when students lose focus is essential to educational data mining systems. The system benefits from timely interventions and personalized learning strategies through this signal. Previous works, however, have not been able to achieve this goal since they only utilize simple ensemble strategies and have shallow feature representations. A meta-learning ensemble architecture with an adaptive mechanism for prediction of student focus is introduced in this paper. It combines residual correction XGBoost, multi-stage gradient boosting with adaptive learning rates, and a confidence-weighted dynamic meta-learner.Our method averts to the major three problems highlighted in the previous work: 1) Single-stage error correction is improved by two-stage residual learning mechanism for the first time; 2) By performance-aware and diversity-enhanced meta-learning, static ensemble weighting is replaced with dynamic ensemble weighting; 3) Four tailor-made behavioral interaction features for student engagement dynamics are used to substitute generic feature engineering. The large-scale dataset of 14,003 students was used for the evaluation of the model, and the results were outstanding: R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>=0.9989 (99.89% variance explained) and RMSE=0.2016. The performance reported here is 17.49% better than the average of existing literatures (82.4%) and 1.89% better than the best single-model approach (98.0%). Moreover, ablation experiments verify that each proposed component leads to quantifiable performance increases. The cross-dataset evaluation on independent Kaggle data resulting in R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>=0.9580 with a 0.41% train-test gap, whereas an 8.0% gap in previous work. SHAP explainability analysis shows that the custom engagement proxy feature is the most influential factor, followed by stress level and attendance.
AI-powered interactive Q&A system for enhanced learning in classrooms R. Umesh, Devi R. Sharmila, R. Keerthana, Sobana Manikandan Artificial Intelligence Computational Intelligence and Inclusive Technologies Proceedings of International Conference on Artificial Intelligence Computational Intelligence and Inclusive Technologies Icraic2it 2025, 2026
AURORA: Hybrid Deep Learning-based Stampede & Risk Prediction System J. Punitha Nicholine, J. Relin Francis Raj, R. Santhana Krishnan, Tamizhselvi Annamalai, Vinoth Kumar V, R. Umesh Proceedings of the 2026 International Conference on AI Driven Smart Systems and Ubiquitous Computing Icauc 2026, 2026 Crowd stampedes in densely populated public spaces pose critical safety risks, requiring accurate and real-time monitoring systems. This research presents AURORA (Advanced Unified Risk Observation & Real-time Alert), a hybrid deep learning framework designed for early crowd stampede risk prediction. AURORA integrates CSRNet-based crowd density estimation, RAFT-lite optical flow motion analysis, Transformer-based spatio-temporal risk modeling, and a rule-based safety layer to ensure explainable and reliable predictions. A custom dataset, AURORA-CROWDX, consisting of annotated video sequences captured at 15 FPS from diverse public environments, supports training, evaluation, and validation. Performance evaluation demonstrates that AURORA outperforms existing methods (CNN-Density and MCNN) across all metrics. For crowd density estimation, it achieves a crowd count accuracy of 0.953, MAE of 0.426, MSE of 0.762, and zone-wise density accuracy of 0.941. Motion analysis results include average crowd speed of 0.812 m/s, speed variance 0.096, direction entropy 0.274, flow consistency 0.941, and panic motion score 0.872. Spatio-temporal risk prediction yields a Transformer risk probability of 0.891, risk prediction accuracy 0.943, and early warning lead time of 13.875 seconds. Hybrid risk assessment reduces false alarms to 0.041 while maintaining classification accuracy of 0.943. System throughput reaches 14.875 FPS with low latency (74.562 ms/frame) and robust performance under occlusion (0.912) and illumination variations (0.931). The proposed framework provides several advantages: high prediction accuracy, early warning capability, explainable decisions, and robust real-time operation. By enhancing crowd safety in public spaces, this research demonstrates a practical and scalable solution for real-time risk management in densely populated urban environments.
Real-World Bug Hunting Techniques for Advanced Real-Time IP Address Redirection R. Umesh, Akchara TD, G. Naglakshmi 6th International Conference on Mobile Computing and Sustainable Informatics Icmcsi 2025 Proceedings, 2025 This research study builds the state-of-art cyber security platform that eventually relied on the internet. It focuses on building this from high-end approaches with practical bug-hunting awareness in its aim toward effective threat detection. It continuously monitors network traffic through several algorithm models for analyzing and stopping malicious activities at real time. In doing this, it blocks and dynamically changes IP addresses in such a way that unwanted access cannot happen. Key Features: Anomaly detection through User and Entity Behavior Analytics, Security Adaptation, Automation, and Response for effective threat management, a threat intelligence feed keeping it abreast of new attack vectors, and the platform comes with deep logging for forensic investigations purposes. Predictive analytics, along with its continuous learning module, allows the detection algorithms to get optimized. Its friendly user interface promotes teamwork, data aggregation, and collective protection and, henceforth, provides the all-encompassing and flexible cybersecurity solution that can help to strengthen the businesses' defenses against constantly changing threats.
Comparative Study of Large Language Models for Adaptive AI Tutoring Systems R. Bastin Jerald, S. Elamaran, R. Umesh 2025 IEEE 7th International Conference on Computing Communication and Automation Iccca 2025, 2025 The study described here is a comparative study of a selection of top Large Language Models (LLMs)—namely Google Gemini, OpenAI's ChatGPT, xAI's Grok, Mistral, and Cerebras—in a dynamic AI-driven tutor system that is capable of adaptive learning. The study used LLMs to assess a learner's knowledge by presenting topic questions that the LLM generates and evaluates automatically, and then delivers a different question difficulty level and/or content type (e.g. blog post, tutorial, video, technical document) based on an automatic evaluation of the learner’s knowledge and abilities. From a learning perspective, all models have capabilities in terms of interaction for educational purposes; however, they each differ in terms of important metrics such as accuracy, reasoning, adaptability, cost-effectiveness, and student independence. Quantitative and qualitative assessments were made in specific technical domains you might be familiar with (e.g. cloud computing, web framework, etc). ChatGPT provides the greatest clarity and reasoning in feedback; Gemini has the most responsive low-latency with the most cost-effective deployment; Mistral provided accuracy and clarity in answers; Cerebras was strong in structured topics and consistency; and Grok was developing in creativity but inconsistent in structured learning flows overall. The study will help to inform the practical trade-offs between LLMs which made for cases in adaptive tutoring systems and the scalability of real-time educational platforms. The integration of explanation tools like SHAP and LIME for greater transparency. This research offers empirical evidence of the adoptive tutoring capabilities of LLMs and provides a roadmap for classroom-based verification in the future.
AI-Powered Offline Voice Assistant for Rural Communities Umesh R, Sudharasan Dev K, Praveen Kumar M Proceedings of the 7th International Conference on Innovative Data Communication Technologies and Application Icidca 2025, 2025 This paper discusses the design and evaluation of an offline, AI-enabled voice assistant that employs Whisper-based speech recognition; lightweight, large language models (LLMs) using Ollama; and pyttsx3-based text to speech (TTS). The offline nature of the voice assistant was designed, in part, to tackle privacy concerns associated with technology that requires a connection to the internet, while still enabling the conversational interactions defined at the beginning of this section. Three LLMs—Gemma 2B, Phi-3, and Mistral—were tested, using a varied set of multilingual and domain-specific contextual prompts (including agriculture, health, and general knowledge, etc.) for their ability to generate appropriate responses. Functional characteristics of the assistant include session memory and a visual modular architecture that enable components and included features to be easily upgraded and interchanged. The Whisper API, in combination with a TTS engine, provided accurate recognition instance-to-instance while TTS produced plausible responses that maintained a natural sense of pacing and development throughout. Performance testing demonstrated between 75% and 90% transcription accuracy in quiet and moderate noise environments, respectively, and an average response time of 2.5 seconds, with significantly lower latency for repeat requests. The fact that the assistant is developed to run on any device with at least 2GB of shared RAM means it will also run on low-resourced or rural technologies. Relative to cloud-hosted alternatives, the proposed voice assistant also has the benefit of privacy, offline reliability, and independence from cloud-based data services and the dangers associated with the widespread, unchecked, use of those services. While the experimental results demonstrate that Mistral model performed the most balanced in terms of response accuracy and clarity (followed by Gemma and Phi-3).
Leveraging Transfer Learning with MobileNetV2 for Amla Leaf Spot Disease Detection in Agricultural Imagery S. Naveen, R. Divya, R. Santhana Krishnan, V. Vinoth Kumar, G. Vinoth Rajkumar, R. Umesh Proceedings of 5th International Conference on Soft Computing for Security Applications Icscsa 2025, 2025 The detection of leaf spot disease in Amla (Indian gooseberry) has been a challenging task due to limited availability of large, high-quality datasets, varied environmental conditions, and the need for real-time monitoring. Traditional methods of disease detection rely heavily on manual inspection, which can be time-consuming, subjective, and prone to errors. Furthermore, the high variability in disease symptoms under different light and climate conditions makes it difficult to deploy a consistent solution across different regions. Current machine learning approaches often struggle with generalization, leading to poor performance in diverse agricultural environments. This work proposes a deep learning-based approach using MobileNetV2, a lightweight convolutional neural network (CNN), to address the challenges in detecting Amla leaf spot disease. The methodology leverages the AmlaLeafSpot-AgriVision and AmlaLeaf-SpotNet datasets for training and validation. Data augmentation techniques such as rotation, flipping, and zooming are applied to enhance the model’s robustness. MobileNetV2 is chosen for its efficiency in handling real-time applications on mobile and edge devices, providing a balance between performance and computational cost. Key features of this work include high accuracy in disease detection, low inference time suitable for deployment in mobile and IoT devices, and scalability for different field conditions. The system also offers automated disease monitoring, reducing the need for manual inspection and allowing for timely intervention in disease management.
Social Media-Based Travel Planning: Integrating Content-Based and Collaborative Filtering for Personalized Suggestions S. Naveen, G. Vinoth Raikumar, R. Santhana Krishnan, R. Umesh, A. Alice Blessie, Kruthika Paulraj Proceedings of the International Conference on Multi Agent Systems for Collaborative Intelligence Icmsci 2025, 2025 This research presents a data-driven approach for providing personalized travel and stay recommendations by integrating social media data., machine learning models., and time-series analysis. Data is collected from social media platforms using APIs., including textual data (captions., hashtags., and comments)., multimedia data (images and videos)., and engagement metrics (likes.,shares., and comments). Textual data is preprocessed using advanced NLP techniques such as BERT for keyword extraction and TextBlob for sentiment analysis., while multimedia data is classified using a pre-trained ResNet model. A comprehensive user travel history is built from geotags and timestamps., and seasonal travel patterns are analyzed using Prophet. User preferences are derived by combining engagement metrics., sentiment scores., and multimedia classification results. Unexplored destinations are identified using K-Means clustering., and similarity scores are computed using cosine similarity between user preferences and destination features. Recommendations for destinations and accommodations are generated using content-based filtering and collaborative filtering with Singular Value Decomposition (SVD). This integrated system ensures context-aware., relevant., and personalized travel and stay suggestions based on user interests and travel history.
