Volume 14 | Issue 7 |

 Abstract

Efficient waste management has become a major challenge in rapidly urbanizing environments due to dynamic and spatially distributed waste generation patterns. Traditional fixed-schedule waste collection systems fail to adapt to real-time conditions, leading to inefficient operations, unnecessary fuel consumption, and frequent overflow scenarios [1]. IoT-enabled smart bin monitoring systems allow continuous real-time data acquisition and improve operational visibility [2]. However, most existing solutions operate reactively and lack predictive decision support for proactive waste collection planning [3].

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Smart Waste Management, IoT, Spatio-Temporal Fusion Transformer, STFT, Machine Learning, Route Optimization, Smart Cities

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

The clinical management of Anterior Cruciate Ligament (ACL) reconstruction recovery is fundamentally dependent on the quality of longitudinal, home-based physical therapy. However, existing diagnostic paradigms suffer from a reliance on subjective patient feedback and intermittent clinical assessments, creating a high risk for graft failure or compensatory gait habits. This paper proposes a novel, high-fidelity Smart Knee Brace that integrates multi-modal sensory inputs including flexible strain sensors, inertial measurement units (MPU-6050), and muscle-kinetic pressure pads. Managed by an ESP32 microcontroller, the system establishes a real-time Internet of Medical Things (IoMT) connection via the Blynk IoT Cloud and a custom-designed web dashboard for clinical telemetry. A Random Forest machine learning classifier is integrated into the software backend to provide autonomous movement classification and safety alerts. We present the comprehensive system architecture, mathematical movement models, and clinical validation across diverse gait patterns. Results indicate a 94.2% accuracy in identifying antalgic gait, positioning the system as a robust tool for enhancing orthopedic tele-rehabilitation and reducing the socio- economic burden of re-injury.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

ACL Rehabilitation, Internet of Medical Things (IoMT), ESP32, Machine Learning, Random Forest, Gait Analysis, Blynk Cloud, Smart Wearables, Flex Sensors, Biofeedback

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Financial fraud has become a major concern in digital transactions due to the rapid growth of online payment systems. Detecting fraudulent transactions in real time is critical to prevent financial losses and enhance security. This paper proposes a real-time fraud detection system using machine learning and streaming technologies. The system integrates Apache Kafka for real-time data streaming, a Random Forest classifier for fraud detection, and SHAP (SHapley Additive exPlanations) for model interpretability. The architecture processes transactions dynamically and classifies them as approved or blocked with minimal latency. A Streamlit dashboard is used for real-time monitoring and visualization. The proposed system improves detection accuracy, reduces response time, and provides explainable insights, making it suitable for modern financial systems.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Explainable Artificial Intelligence, Fraud Detection, Model Transparency, Stream Processing, Payment Fraud, SHAP, Real-time Analytics.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

As healthcare systems become more digitized, there has been a significant increase in electronic medical records, diagnoses, and clinical documents. Yet, the specialized nature of the language used in many reports makes it hard for non-medically trained people to understand them fully, resulting in poor health literacy and possibly causing erroneous interpretations. In this paper, we propose MediScan AI, which is an artificial intelligence platform for automatic medical report interpretation using explainable approaches. Specifically, MediScan AI incorporates Optical Character Recognition (OCR), Biomedical Named Entity Recognition (NER), transformer-based summarization, and knowledge retrieval into a single pipeline. Tesseract is used for OCR, and SciSpacy & BioBERT are used for NER tasks. For summarization purposes, transformer models such as DistilBART and FLAN-T5 are leveraged. Knowledge retrieval is accomplished by using the MedQuAD dataset, together with TF-IDF & BM25 as ranking algorithms. The experimental results show an F1-score of 0.93 for the biomedical NER task and a competitive ROUGE score compared to baselines.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Explainable AI, Medical Report Analysis, Biomedical NLP, SciSpacy, BioBERT, Transformer Models, DistilBART, FLAN-T5, Tesseract OCR, Knowledge Retrieval, TF-IDF, BM25, MedQuAD

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Modern IT environments are highly complex and technical, and prone to sudden system failures that impact productivity and significantly increase the cost of maintenance. Traditional fault detection methods rely on reactive support ticket generation, where issues are addressed only after system failure. This paper presents a machine learning-based framework that proactively identifies faults using real-time system performance metrics. System parameters including CPU utilization, memory usage, disk activity, network latency, application response time, and security indicators are continuously monitored.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Predictive Maintenance, Machine Learning, Failure Classification, Support Ticket Generation, Random Forest, KNN, SVM, System Monitoring

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Automated credit risk assessment in modern banking institutions must simultaneously serve two structurally distinct borrower populations: individual retail applicants and corporate entities, whose financial profiles, data representations, and risk drivers differ fundamentally. Existing systems either employ a single unified model that sacrifices domain specificity, or maintain separate models without a coherent architectural framework for integrating them into a consistent decision pipeline. This paper presents CreditAI, a dual-domain framework deploying two independently designed LightGBM classifiers within a shared microservice architecture while preserving the domain-specific feature engineering and probability-to-score conversion logic required for each segment. The retail model processes a 10-feature vector encoding income capacity, debt burden, and payment history via a log-odds scoring function. The corporate model processes a 15-feature vector of accounting ratios, including five composite interaction terms encoding cash flow stress and leverage risk, via a linear scoring function. A transparent three-tier decision engine routes scored applications to automated approval, automated rejection, or mandatory officer review. Finalised decisions are cryptographically committed to an Ethereum smart contract for tamper-proof regulatory auditability. Evaluation on public benchmark datasets yields AUC-ROC scores of 0.941 and 0.913 for the retail and corporate models respectively, outperforming unified- model and domain-agnostic baselines by margins of 2.1 to 4.4 percentage points.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

LightGBM, Dual-Domain Framework, Hyperparameter Optimisation, Domain-Specific Feature Engineering, Corporate Bankruptcy Prediction, Automated Decision Engine, Blockchain Audit Trail, Multilingual Voice support

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

The rapid growth of online fashion platforms has increased the demand for personalized and intelligent recommendation systems. However, existing fashion recommendation approaches often rely solely on user history or collaborative filtering techniques, resulting in generic suggestions that fail to consider individual physical attributes, contextual preferences, and styling requirements. Moreover, most systems lack the ability to integrate visual cues such as body shape, skin tone, and fabric patterns, which are essential for effective fashion personalization. This paper proposes a hybrid multi-modal fashion recommendation system that combines machine learning, computer vision, and rule-based fashion knowledge to generate highly personalized outfit and styling suggestions. The system accepts both manual inputs and visual inputs to extract user-specific attributes including body shape, skin tone, and facial features. A category-aware framework is employed to ensure recommendations are constrained within user-selected clothing types across both men's and women's fashion domains.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Multi-modal fashion recommendation system, Content based filtering, Machine Learning algorithm.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Current computer networks function in cloud computing environments which undergo constant changes because users and devices and services keep connecting to different systems throughout the network. The existing security methods depend on fixed access control rules which determine trust based only on the first authentication and maintain trust until the user session ends. The system creates essential security weaknesses because it allows inside attackers and credential theft and lateral movement to occur when it does not track system operations in real time. The research introduces an Adaptive Behaviour-Driven Zero Trust Control Plane which establishes continuous protection for network security operations. The system records all TCP and UDP network data which it converts into flow patterns through specific time intervals that generate network usage data which includes total packets and total bytes and total session time. An unsupervised Isolation Forest model is employed to detect anomalies without requiring labelled data, enabling the identification of unknown and emerging threats. The system uses statistical baselines to assess current trust levels through dynamic analysis of generated anomaly scores in order to determine present system status. The system uses trust levels to control an adaptive policy engine which implements rate limiting and access blocking and real-time alerting functions. The framework functions as a modular system which connects traffic acquisition and anomaly detection and trust evaluation and policy enforcement and visualization to create a cloud-based system that offers real-time monitoring at scale. The testing results show that the system can detect anomalies successfully while applying security measures which boost the security strength and quick response ability of contemporary network protection technologies.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Zero Trust Architecture, Anomaly Detection, Isolation Forest, Network Security, Trust Computation, Real-Time Monitoring, Adaptive Security.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

The retail industry generates large volumes of sales and customer data, making accurate demand forecasting essential for efficient inventory management and business growth.Traditional forecasting methods rely on static models and historical averages, which often fail to capture changing customer behavior and market trends.This project proposes a Smart Retail Insights system that leverages Artificial Intelligence and Machine Learning techniques to predict store sales and forecast product demand.The system analyzes historical sales data, seasonal trends, promotions, and customer purchasing behavior to generate accurate demand predictions.A key feature is Dynamic Customer Lifetime Value (CLV), which continuously evaluates customer value based on recency, frequency, and spending patterns.The proposed framework helps retailers optimize inventory planning, reduce stock shortages,

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Artificial Intelligence, Machine Learning, Retail Analytics, Customer Churn Prediction, Customer Lifetime Value, Demand Forecasting, Inventory Management, Natural Language Processing, Decision Support System.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

The rapid growth of customer transaction data has created a need for intelligent systems capable of analyzing behavior and predicting product trends effectively. This paper proposes an Explainable AI-based framework that integrates feature engineering, deep learning, clustering, and decision-making techniques to analyze customer behavior and forecast trends. The system utilizes Recency, Frequency, and Monetary (RFM) analysis along with derived features such as purchase intervals and seasonal attributes to represent customer activity. A Multilayer Perceptron (MLP) model is employed to predict trend probability, while K-Means clustering is used for customer segmentation. To enhance transparency, Explainable AI techniques are incorporated to interpret model predictions, and a rule-based decision engine converts outputs into actionable strategies. Additionally, counterfactual analysis enables "what-if" scenario evaluation, and trend and seasonal analysis identify demand patterns over time. Experimental results demonstrate that the proposed approach achieves a prediction accuracy of approximately 96.5%, along with effective segmentation and improved interpretability, making it a valuable tool for data-driven business decision-making.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Product Trend Prediction, K-Means Clustering, Multi-Layer Perception, RFM Analysis, Customer Segmentation, Explainable AI, Seasonal Demand.

  License

Creative Commons Attribution 4.0 and The Open Definition