VOLUME 15, ISSUE 8, AUGUST 2026

Abstract: AI agents are being used in construction collaboration platforms increasingly to perform tasks related to planning, safety, subcontracting, document validation, and inspection among others. However, highly sensitive data such as contractual details, finances, personnel information, and design layouts are processed through collaboration platforms. In order to facilitate coordinated functions among different AI agents without compromising their integrity, security, and regulatory compliance, this paper proposes a Secure AI Orchestration Framework (SAIOF) that would support collaboration platforms development. The SAIOF framework encompasses role-based access controls, multi-agent coordination through policies, agent orchestration federations, and encrypted communication among AI agents. The proposed model will allow inspectors, subcontractors, and project managers to collaborate securely without unnecessarily exposing sensitive or confidential data. As can be seen from the simulation experiment of a construction collaboration workflow, there is less risk of data leaks, greater efficiency, and greater workflow transparency with this orchestration approach. From the analysis of experimental results, it follows that secure orchestration improves the accuracy of coordination actions, limits malicious attempts to gain unauthorized access to sensitive information, and increases scalability in the deployment of AI systems in smart buildings.

Keywords: AI orchestration, intelligent coordination, safe multi-agent systems, federated AI, construction cybersecurity, access control, and privacy-preserving AI

Abstract: Existing IoT Intrusion Detection Systems (IDS) achieve competitive accuracy but are opaque black-box models, relying on single imbalanced datasets, centralising raw traffic, and incompatible with edge hardware. This paper presents XAI-IDS, a seven-layer federated explainable IDS resolving all four limitations simultaneously. The system trains LightGBM and a CNN-LSTM hybrid on CICIOT2023 (80/10/10 split, 5-fold CV) and evaluates without retraining on ToN-IoT and NSL-KDD across 33 attack types. LightGBM achieves macro-F1=0.9412/0.9138/0.9267 on CICIOT2023/ToN-IoT/NSL-KDD; CNN-LSTM achieves 0.9534. FPR=3.12% vs 8.92% (Extra Trees [4], p<0.001, - 5.80 p.p.). SMOTE-ENN recovers normal-class F1 from 0.6312 to 0.9134. TreeSHAP provides exact attribution in 4.3±0.4 ms; LIME cross-validates at 83.2% top-5 agreement; feature ablation confirms faithfulness (ΔAcc=19.3±1.7 p.p.). An LLM explanation layer raises non-technical comprehension from 31.2% to 84.3% (t(19)=14.23, p<0.001, n=20, ethics-approved). FedProx μ=0.01 achieves F1 within 1.32% of centralised baseline in 7 rounds under DP-SGD ε=5. ADWIN recovers from 93.7% of drift events in 47.3±4.8 s. The ONNX INT8 LightGBM (14.2 MB) achieves 1.83 ms inference and +0.4 W power draw on Raspberry Pi 4. All improvements: p<0.001, Bonferroni-corrected. Complete hyperparameters (Table IX), 78-feature schema (Table X), and edge metrics (Table XI) provided for full reproducibility.

Keywords: IoT Security, IDS, XAI, SHAP, LIME, LightGBM, CNN-LSTM, FedProx, Differential Privacy, SMOTE- ENN, ADWIN, Concept Drift, ONNX, LLM.

Abstract: Creating multilingual voice content today typically requires re-recording narration in every target language or stitching together separate tools for transcription, translation, and speech synthesis. This paper presents AuraVoice, an end-to-end web-based platform that unifies voice cloning, cross-lingual speech translation, and automated video dubbing in a single pipeline. The system combines OpenVoice V2 for zero-shot tone-color conversion, faster-whisper for automatic speech recognition, MeloTTS for multilingual base speech synthesis, and FFmpeg for audio/video processing, orchestrated through a FastAPI backend with a lightweight browser-based frontend. Given a short reference audio sample, AuraVoice extracts speaker tone characteristics and reproduces them in synthesized speech across six languages while preserving the original speaker's identity. The video dubbing module extends this pipeline to full videos by extracting the source audio, translating and re-synthesizing it in the target language, and re-muxing it with the original visual track. We describe the system architecture, module-level implementation, and lazy-loading/caching strategy used to manage GPU memory across multiple deep learning models, and discuss the functional testing performed on each module. The platform demonstrates that recent advances in zero-shot voice cloning and open automatic speech recognition can be composed into a practical, integrated tool for multilingual content creation, reducing the manual effort required compared to disjoint single-purpose tools.

Keywords: Voice Cloning, Tone-Color Conversion, OpenVoice, Text-to-Speech, Speech Translation, Video Dubbing, Whisper, FastAPI, Speech Synthesis

Abstract: SQL Injection attacks are one of the most critical and frequently occurring security threats in modern web applications. These attacks exploit vulnerabilities in input handling to inject malicious SQL statements, enabling attackers to bypass authentication, retrieve confidential data, modify database contents, or even gain complete control over backend systems. Traditional defense mechanisms such as manual input validation, rule-based filtering, and signature-based intrusion detection systems are limited in their ability to detect complex, obfuscated, and previously unseen attack patterns. This project proposes a Smart SQL Injection Attack Detection System using a Hybrid Deep Neural Network (CNN + LSTM) to effectively identify both known and unknown SQL injection attacks. Incoming SQL queries are preprocessed through tokenization and encoding, converted to numerical form using TF-IDF, and refined with Chi-Square feature selection. CNN layers capture local attack signatures while LSTM layers model sequential and contextual relationships among query tokens, jointly improving learning capability and detection accuracy. The model is trained and evaluated on benchmark and real-world SQL injection datasets using accuracy, precision, recall, and F1-score. Experimental results show detection accuracy consistently above 95%, with lower false positive rates than traditional and single-model detection techniques, and the system is suitable for real-time deployment in existing web application security infrastructures.

Keywords: SQL Injection, Cybersecurity, Deep Learning, Hybrid Neural Network, CNN, LSTM, Web Security, Attack Detection.

Abstract: This paper presents Mitro, a desktop human-computer interaction (HCI) system that fuses three input modalities into a single interaction pipeline: (1) computer-vision-based hand-gesture recognition for direct manipulation of the operating system (cursor control, clicks, scroll, volume, and brightness); (2) speech-based natural-language commanding through automatic speech recognition (ASR) and text-to-speech (TTS); and (3) a large language model (LLM) reasoning layer that replaces brittle keyword-matched command dispatch with open-domain natural-language understanding and tool/function calling. Mitro is built as an upgrade to an existing rule-based gesture-and-voice controller that suffered from three concrete limitations: order-sensitive substring-matched command routing, a single hardcoded audio energy threshold that did not generalize across acoustic environments, and a single hardcoded gesture-classification threshold with a fragile single-frame-reset debounce scheme. This work addresses all three limitations with a unified calibrate-once-then-lock design pattern applied consistently across the audio and vision pipelines, and generalizes command handling from rule-based matching to LLM-mediated tool dispatch while preserving deterministic, network- independent fast paths for safety-critical control flow. Gesture-state changes, voice transcripts, and LLM responses are surfaced into a single chronological interaction log, providing an observable, demonstrable trace of the system's multimodal behavior. Section IV outlines the evaluation methodology used to measure calibration benefit, gesture- smoothing benefit, command-routing robustness, and end-to-end latency, together with the qualitative findings obtained during development; the corresponding quantitative results are to be reported once the described experiments are executed on the deployed system.

Keywords: Multimodal Human-Computer Interaction, Gesture Recognition, Speech Recognition, Large Language Models, Tool Calling, Adaptive Calibration, Temporal Smoothing, Cross-Modal Fusion, MediaPipe, Groq

Abstract: Organizations increasingly rely on cloud platforms to store data, run applications, and manage IT infrastructure, drawn by the flexibility and on-demand scalability these platforms offer. That flexibility comes at a price, however: keeping cloud spending under control is one of the more persistent operational headaches for teams running workloads at scale. Resources sit over-provisioned or idle far more often than they should, and the resulting waste translates directly into avoidable expense. Bringing smarter, more automated cost control into everyday cloud operations is no longer optional as usage keeps climbing. This paper introduces CloudOpt-AI, a predictive FinOps framework that pairs machine learning with multi-cloud resource scheduling to tackle exactly that problem. The framework studies past usage patterns to project near-term resource needs and then allocates capacity accordingly. It also coordinates workload placement across several cloud providers at once, so that organizations can pick cost-effective resources without sacrificing performance or reliability. Demand forecasting, automated allocation, and FinOps discipline are woven together so that CloudOpt-AI supports faster, better-informed decisions and squeezes more value out of existing cloud capacity. The overarching goal is threefold: cut operational spend, raise resource efficiency, and make cloud cost governance genuinely manageable rather than a manual chore. Taken together, the results point to how predictive analytics and intelligent scheduling can help organizations strike a workable balance between cost and performance in today's cloud environments.

Keywords: Cloud Computing, Cloud Cost Optimization, FinOps, Machine Learning, Multi-Cloud Computing, Resource Allocation, Resource Scheduling, Demand Forecasting, Cost-Aware Computing, Workload Management.

Abstract: Depression remains one of the most widespread mental health conditions worldwide, and social media has become an unusually rich window into how people express emotional and behavioral change. Most existing detection systems however still look at a single channel either the wording of a post or a single accompanying image and largely ignore how a person’s posting behavior shifts over time. This paper addresses that gap with a temporal, multi-modal pipeline that reads depression risk not from one post in isolation but from a sequence of a user’s posts, combining what is written, what is pictured, the emotional tone of both, and how these change across time. Images are converted into descriptive captions using BLIP, which are then encoded alongside raw visual features and a set of behavioral and temporal signals processed through a GRU. When these three streams text, image, and temporal-behavioral are fused, the resulting model reaches 71% classification accuracy. The results support the view that behavioral and temporal cues, even when individually weak predictors, meaningfully sharpen a multi-modal system’s ability to separate depressed from non-depressed posting patterns.

Keywords: Depression Detection; Social Media Analytics; Multi-modal Fusion; Temporal Behaviour Modelling; Image Captioning; Sentiment Analysis; BERT; GRU.

Abstract: Blood is a biologically-complex two-phase fluid consisting of suspended cellular (blood) components and plasma. Blood's properties of transport affect the function of the heart while pumping under pulsatile flow conditions. A general and total analytical model for plasma transport and blood-cell transport through a cylindrical artery experiencing pulsatile pressure drops is developed. Equations used for developing the model are the incompressible Navier-Stokes equations with viscosity depending on hematocrit (the volume fraction of red blood cells), a convective- diffusive-migrative equation describing cellular transport, and an advective-diffusive equation describing plasma solute transport. Analytical solutions for the parameters are determined by separating variables and solving them using the axisymmetric fully developed pulsatile flow analytical solution of Womersley. In addition to giving the spatial distributions of velocity, hematocrit, plasma concentration, cellular concentration, wall shear stress, and volumetric flow rate the solutions provide information concerning changes that occur due to varying hematocrit levels. Numerical simulation results indicate that higher hematocrit values result in increased effective viscosity of blood. Increased blood viscosity causes decreased velocities in the direction of flow, decreases the maximum flow rates occurring during the pulse cycle, and increases frictional resistance to flow. Additionally, numerical simulation results demonstrate the existence of a "cell-rich core" along the centerline of vessels and a "plasma-depleted zone" near vessel walls. These zones develop because shear-induced migration of red blood cells towards the centerline. Plasma concentration decreases continuously downstream from the inlet due to a combination of advection, diffusion, and wall transport.

Keywords: Pulsatile blood flow, Plasma transport, Cellular transport, Hematocrit, Mathematical modeling, Non- Newtonian blood

Abstract: The rapid growth of digital technology has transformed the way people manage their health and fitness. While many online platforms allow users to purchase fitness supplements, they often fail to provide personalized guidance based on an individual's health condition and fitness goals. This project, AI Fitness and E-Commerce Platform (POP GYM), addresses this gap by combining an online supplement store with an AI powered diet recommendation system.The platform enables users to create an account, browse and purchase fitness supplements, manage their shopping cart, and receive customized diet plans based on personal details such as age, height, weight, gender, and fitness objectives. The diet recommendations are generated using Google's Gemini AI, making them more relevant and personalized for each user. In addition to user features, the system includes an admin dashboard that allows efficient management of products, users, and customer orders.The application is developed using React.js for the frontend, Node.js and Express.js for the backend, and MySQL for database management. This combination ensures a responsive, secure, and scalable web application. By integrating e-commerce functionality with AI-driven health recommendations, the platform offers users a convenient and intelligent solution to support both their fitness shopping and nutritional planning.The project demonstrates how Artificial Intelligence can enhance user experience by providing personalized recommendations while simplifying the management of fitness-related products and services. It also provides a strong foundation for future enhancements such as online payment integration, workout tracking, wearable device connectivity, and mobile application support.

Keywords: Artificial Intelligence, Fitness Management, E-Commerce Platform, Personalized Diet Recommendation, Google Gemini API, React.js, Node.js, Express.js, MySQL, Web Application, Health and Nutrition, Supplement Management, User Authentication, Admin Dashboard, Responsive Web Development.

Abstract: Imagine what your smartphone already knows about you based on how you move. Each step, each climb up the stairs, each slouch into the chair – it is all recorded by your smartphone sensors. To unlock the immense possibilities of using this data for the identification of human actions in real-time, while keeping your privacy intact is an extremely promising endeavour . In this paper, EE-FL-HAR is presented, a system that allows training activity recognition models right inside the smartphone, without transferring any sensor data out of the device itself. Only the model updates are exchanged between devices through the network. The results were tested on a 2,004 sample dataset with 695 features and covering six different daily activities. The Random Forest model achieved 83.29% accuracy with the precision of 0.8348, recall of 0.8329 and F1-score of 0.8326. Federated learning simulation of 50 rounds approaches 93% of the accuracy of a centralized baseline, reducing the total energy consumption by 56%. Six diagnostic figures show what is happening in every step. The conclusion is clear: activity recognition and privacy do not need to be opposing forces anymore.

Keywords: Federated Learning, Human Activity Recognition, Sensor Suite of Smartphone, Energy Efficiency, Personalization, Privacy-preserving Machine Learning, Random Forest, Communication Efficiency.

Abstract: Modern-day education has changed due to the utilization of artificial intelligence by implementing adaptive learning systems tailored towards individual students. Traditional learning methods generally present the same content to every learner without considering individual learning speed, understanding, or preferences. To address this limitation, this project proposes Data Spark Assistant – An Agentic AI Powered Multimodal RAG Learning System, an intelligent learning platform that delivers personalised educational support using Artificial Intelligence (AI). The system combines Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), Natural Language Processing (NLP), and Agentic AI to analyse learner performance and generate customised study materials. Through intelligent quizzes and assessments, the platform identifies each learner's strengths, weaknesses, and knowledge gaps. Based on this analysis, it recommends personalised lessons, practice questions, and learning resources that help students improve in weaker areas. By integrating adaptive learning techniques with AI-driven content generation, the proposed system creates an interactive and personalised learning environment. This approach enhances learning efficiency, encourages self-paced study, promotes critical thinking, and supports better academic performance.

Keywords: Adaptive Learning, Agentic AI, Retrieval-Augmented Generation (RAG), Large Language Models (LLMs), Natural Language Processing (NLP), Knowledge Gap Analysis, User Profiling, Multimodal Learning.

Abstract: Nowadays, sharing important files over the internet has become very common, but it also increases the risk of unauthorized access and data theft. Many existing file-sharing systems have security issues such as weak authentication and poor access control. This paper presents a Secure File Sharing System with Multi-Factor Authentication (MFA) to improve the security of file sharing. The system is developed using Next.js, React, Better- SQLite3, Bcrypt, JSON Web Token (JWT), and Time-based One-Time Password (TOTP). It allows users to securely upload, share, and download files after successful authentication. Passwords are securely stored using Bcrypt, while JWT provides secure user sessions. TOTP-based Two-Factor Authentication adds an extra layer of security to prevent unauthorized access. The proposed system is easy to use, provides secure file sharing, and offers better protection against common cyber threats.

Keywords: Secure File Sharing, Multi-Factor Authentication, JWT, Bcrypt, TOTP, Next.js, Better-SQLite3, Web Security

Abstract: Agricultural planning under variable soil and climatic conditions requires simultaneous consideration of productivity, profitability and limited irrigation resources. This paper presents the Neuro-EvoSwarm Optimizer (NESO), a hybrid decision-support framework that integrates Deep Learning (DL), Genetic Algorithm (GA), and Particle Swarm Optimization (PSO) for seasonal multicropping strategy optimization under drip irrigation. The framework uses soil characteristics, climatic variables and crop-management information to predict crop yield and economic profit and then searches for high-quality crop combinations. Data preprocessing includes missing-value handling, outlier detection, categorical encoding, feature scaling and normalization. The Deep Learning component learns nonlinear relationships between agricultural conditions and target outcomes. GA provides broad evolutionary exploration through population initialization, fitness evaluation, selection, crossover, mutation and replacement, while PSO refines the best GA solutions using personal-best and global-best information. A multi-objective fitness formulation is used to maximize yield and profit while minimizing irrigation water use. The source document describes the model architecture, workflow, training configuration and decision-support outputs, but does not provide a complete set of numerical test results; therefore, this condensed paper does not fabricate accuracy or optimization values. The framework provides a scalable basis for precision agriculture and can be extended with real-time IoT observations, weather forecasts, GIS and additional sustainability indicators.

Keywords: seasonal multicropping; drip irrigation; Deep Learning; Genetic Algorithm; Particle Swarm Optimization; NESO; crop yield prediction; agricultural decision support; water-use efficiency

Abstract: Automatic Number Plate Recognition (ANPR) is an important computer vision technology used in Intelligent Transportation Systems (ITS) for traffic monitoring, law enforcement, parking management, toll collection, and vehicle access control. Previous ANPR systems have primarily relied on conventional image processing techniques, including thresholding, edge detection, morphological operations, and rule-based plate localization. Although these methods can perform effectively under controlled conditions, their accuracy is often affected by variations in illumination, vehicle orientation, background complexity, plate size, and image quality. To address these limitations, this proposed research presents an enhanced Automatic Number Plate Recognition system that integrates the YOLOv9(You Look Only Once Version 9) object detection algorithm with image processing and Optical Character Recognition (OCR) techniques. YOLOv9 is used to automatically detect and accurately localize vehicle number plates from input images, eliminating the need for predefined assumptions regarding the position and size of the plate. The detected plate region is then processed using grayscale conversion, noise reduction, histogram equalization, morphological operations, and Sobel edge detection to enhance the image and improve character separation. The segmented alphanumeric characters are subsequently recognized using an OCR module, and the extracted registration number can be matched with a database for vehicle information retrieval. By combining deep learning-based object detection (YOLO) with median filter and OCR, the proposed approach improves the robustness and reliability of the ANPR system under varying image conditions. To test the hypothesis, this research utilized the publicly available zenodo dataset for the vehicle images used to train, validate, and test the Automatic Number Plate Recognition (ANPR) system. It provides the information required for the YOLOv9 model to learn how to detect number plates and for the OCR module to recognize the characters correctly. It is a diverse dataset containing different vehicle types, lighting conditions, viewing angles, and backgrounds helps improve the accuracy and robustness of the proposed system. Experimental results demonstrate that the proposed system achieves 94.1% number plate detection accuracy and 97.2% Character recognition accuracy, indicating significant enhancement over previously developed ANPR techniques and demonstrating the effectiveness of the integrated YOLO- based ANPR framework for practical vehicle identification applications.

Keywords: Image Processing, Text Extraction, Character recognition, YOLOv9, ANPR, Number Plate Detection, MATLAB