International Journal of Intelligent Computing Systems

Conversational artificial intelligence has shifted decisively from rule-based dialogue management toward neural architectures capable of cross-domain, context-sensitive responses. This paper presents the design, implementation, and evaluation of a transformer-based conversational agent built on a medium-sized pretrained model, integrated through the Hugging Face Transformers library. The system addresses three persistent limitations of conventional chatbot systems: shallow contextual understanding, brittle handling of unanticipated input, and an inability to sustain multi-turn dialogue. We describe a complete processing pipeline encompassing tokenization, autoregressive decoding under the attention mechanism, conversational history tracking, and response refinement, deployed behind a Gradio-based interface for low-friction human evaluation. The architecture further incorporates a fallback layer that intercepts ambiguous or out-of-distribution queries before they degrade the user experience. Qualitative evaluation across representative dialogue sessions indicates that the system reliably preserves topical continuity across turns and produces responses that are syntactically well-formed and pragmatically appropriate, while also revealing characteristic weaknesses of single-domain language models — including occasional repetition and generic responses — that are discussed in relation to systems such as Meena and BlenderBot. The findings substantiate the practical viability of pre-trained transformer architectures as the foundation for scalable, domain-adaptable conversational agents, and motivate a concrete agenda for fine-tuning, multilingual extension, and affect-aware response generation in future iterations of the system.

Driver drowsiness is responsible for a significant proportion of road traffic fatalities worldwide, with the National Highway Traffic Safety Administration (NHTSA) attributing approximately 2.5% of all fatal crashes to drowsy driving. Conventional detection approaches relying on physiological sensors, steering dynamics, or vehicle-lane monitoring are either intrusive, expensive, or impractical for everyday commuters. This paper proposes a real-time drowsiness detection system based on a custom Convolutional Neural Network (CNN) trained on the YawDD benchmark dataset. The system analyses two principal behavioral cues: Eye Aspect Ratio (EAR) computed from DLIB's 68-point facial landmark predictor, and Mouth Aspect Ratio (MAR) for yawning detection, extracted continuously from a front-facing camera feed. The proposed DrowsinessDetector CNN achieves a test accuracy of 99%, precision of 99.47%, and recall of 98.85%, outperforming state-of-the-art methods including VGG16 (~94%), ResNet50V2 (~95%), and InceptionV3 (~96%). Gradient-weighted Class Activation Mapping (Grad-CAM) confirms that the model correctly attends to periocular and perioral regions. The solution is integrated into an Android application (DRIVEGUARD) featuring accelerometer-based crash detection and GPS location tracking, making it suitable for real-world deployment without specialized hardware.

Advanced Persistent Threats (APTs) represent the most sophisticated and strategically consequential class of cyber adversaries, characterized by prolonged network intrusion, multi-stage attack orchestration, and deliberate evasion of conventional detection mechanisms. The attribution of APT campaigns to specific threat actor organizations is a critical yet methodologically challenging undertaking, owing to the volume, heterogeneity, and structural complexity of the underlying cybersecurity data — spanning threat actor profiles, malware families, indicators of compromise (IOCs), exploited vulnerabilities, and adversarial attack campaigns. This paper presents CSKG4APT, a Cybersecurity Knowledge Graph framework specifically engineered for Advanced Persistent Threat organization attribution. The proposed system ingests cyber threat intelligence from multiple authoritative sources and constructs a semantically enriched knowledge graph in which cybersecurity entities are represented as typed nodes and their interrelationships as directed edges, enabling structured reasoning over the threat landscape. To enhance threat classification and attribution accuracy, a hybrid deep learning architecture integrating Gated Recurrent Unit (GRU) networks with Bidirectional Long Short-Term Memory (Bi-LSTM) layers is employed, enabling the model to capture both temporal dependencies and bidirectional contextual patterns within sequential attack event data. Random Forest and XGBoost ensemble algorithms serve as interpretable machine learning baselines for comparative performance analysis. Experimental evaluation demonstrates that the Bi-LSTM–GRU model achieves 99% attribution accuracy, outperforming XGBoost (97%) and Random Forest (96%) baselines.

Phishing is a widely used cyber-attack technique in which users are deceived into visiting illegitimate websites that closely resemble legitimate ones. These fake websites are designed to trick users into revealing sensitive information such as usernames, passwords, bank details, and credit card information. Due to the growing sophistication of such attacks, phishing has become a serious security concern. In the proposed method, we focus solely on analyzing the URL of a website to determine whether it is a phishing site, thereby eliminating the need to visit the website and risk exposure to malicious code. This approach enhances user safety and reduces the chances of infection from harmful scripts or malware embedded in phishing pages. We explore how metadata extracted from URLs — such as domain age, presence of special characters, URL length, and redirection patterns — can help in identifying phishing attempts. These features are used to train a machine learning model using the Random Forest algorithm, known for its robustness and ability to handle high-dimensional data without overfitting.

The exponential growth of user-generated video platforms such as YouTube has transformed the way people communicate, share opinions, and consume digital content. Among the various interaction mechanisms provided by YouTube, the comment section plays a crucial role in facilitating engagement between creators and viewers. However, the openness of this feature has also led to a significant rise in spam comments, including promotional advertisements, phishing links, misleading information, and bot-generated messages. This paper proposes a comprehensive supervised machine learning-based framework for automatically detecting and filtering spam comments in YouTube comment sections. Multiple classification algorithms, including Logistic Regression, Support Vector Machine (SVM), Decision Tree, and Artificial Neural Network (ANN), are implemented and evaluated. Experimental results demonstrate that SVM outperforms other models in terms of accuracy, precision, and F1-score.

This study examines the mechanical performance of a connecting rod designed for a 150 cc single-cylinder four-stroke internal combustion engine. The connecting rod experiences complex alternating compressive and tensile stresses during engine operation. Using finite element analysis, we evaluated both static structural behavior and fatigue characteristics of this critical component, modeled in CREO Parametric and simulated in ANSYS Workbench. We compared conventional forged steel against Aluminium Alloy 7075-T651 under identical loading conditions. Forged steel experienced less deformation (0.001885 mm) but had higher peak stress at 49.748 MPa. In contrast, Aluminium Alloy 7075 showed greater deformation (0.003640 mm) but significantly lower stress at 33.602 MPa, along with a better fatigue safety factor. The most significant advantage of Aluminium Alloy 7075 is its weight — approximately 64% lighter than forged steel — making it a viable lightweight alternative for small-displacement engine connecting rods.

Social media has become an integral part of student life, influencing communication, collaboration, and learning behaviors. This study examines the impact of Social Media Usage (SMU) on Student Academic Performance (SAP), with a focus on how demographic factors such as gender, department, and program type shape social media behavior. A structured questionnaire was distributed to 150 students, resulting in 80 validated responses used for analysis. Correlation analysis, ANOVA, and linear regression were combined with machine learning methods to improve predictive accuracy and reveal hidden patterns. The findings reveal that SMU has a significant and measurable impact on SAP, with AI-driven dashboards effectively identifying students at risk of underperformance and offering educators a visual tool to design targeted interventions.

Floods continue to be one of the most destructive natural disasters, disrupting millions of lives and causing significant economic losses annually. Traditional hydrological and statistical models struggle to represent the nonlinear relationships between rainfall, soil moisture, and river discharge. This paper proposes a machine learning-based flood forecasting framework using SVM, KNN, Logistic Regression, and Multi-Layer Perceptron. MLP demonstrated superior predictive performance due to its ability to capture complex nonlinear dependencies, achieving higher accuracy and faster response time than traditional models.

This paper presents an adaptive traffic light controller based on fuzzy logic for improving traffic flow at an isolated intersection. A set of fuzzy rules uses collected information from road detectors — queue length, arrival flow, and exit flow — to compute the amount of time by which the next phase should be shortened or extended. The proposed fuzzy control system is constituted of two parts: one for the primary driveway with a higher volume of vehicles, and one for the secondary driveway with a lower volume.

Reviews are the most crucial opinions delivered by individuals based on their experiences with a product or service. This study explores individual perspectives on emotional tendencies by employing a deep learning-based text mining approach on the Disneyland reviews dataset, using the Text Blob package to extract polarities of text, phrases, or words. A sophisticated K-Fold Cross Validated LSTM model classifies sentiments and provides predictive analysis, resulting in an average accuracy of 0.9435 and F1-score of 0.936 — outperforming traditional machine learning methods.

Diabetes mellitus represents a global health crisis, affecting over 537 million adults in 2021, with projections reaching 783 million by 2045, often accompanied by comorbidities that escalate mortality risks by 2–4 times. This paper introduces an ensemble machine learning framework integrating Random Forest, XGBoost, and Logistic Regression applied to the PIMA Indians Diabetes Dataset and augmented clinical records, achieving 98.07% accuracy, 97% AUROC, and 96% F1-score. Deployable on modest hardware, the system enables real-time clinical decision support, reducing false negatives by 18% and facilitating proactive interventions.

Credit card fraud detection continues to be a major challenge in the financial industry due to extreme class imbalance, where fraudulent transactions occur far less frequently than legitimate ones. This paper introduces a fraud detection framework that utilizes Variational Autoencoders (VAEs) for generating synthetic data and SMOTE to balance the minority class. We evaluated multiple classification models, including XGBoost, Deep Neural Networks, AdaBoost, and CatBoost, using an augmented dataset. Extensive experiments demonstrate that our hybrid augmentation strategy significantly enhances fraud detection performance by increasing recall and F1-score while reducing false positives.

Small object detection in remote sensing images is essential for urban planning, environmental monitoring, disaster relief operations, and defence. Conventional object detection models, such as Faster R-CNN and RetinaNet, have difficulty preserving fine-grained spatial information. This paper applies YOLOv8 to improve detection of small objects, trained and tested on the DIOR dataset. Experimental results show that YOLOv8 obtains a mAP@50 of 73.4% on training and 71.0% on testing, outperforming previous YOLO versions and conventional detectors, and providing the best speed-accuracy trade-off for real-time applications compared with transformer-based models such as DETR.

Text-to-image generation has seen remarkable progress with the emergence of deep learning models like Stable Diffusion, but conventional approaches often require extensive computational resources and subject-specific fine-tuning. Instant Imager eliminates this need by leveraging in-context learning, enabling the model to replicate the capabilities of numerous subject-specific expert models. The framework generates customized images 10 times faster than conventional optimization methods, while delivering a user-specific number of images with superior quality, streamlining the image generation process for artists, designers, and content creators.

Object detection plays a crucial role in developing intelligent and efficient traffic management systems. This paper introduces a real-time object detection system powered by YOLOv8 that can identify and classify various vehicle types — cars, trucks, buses, motorcycles, and bicycles — from live traffic surveillance footage. Experimental results indicate that YOLOv8 delivers high detection accuracy alongside real-time processing, making it highly viable for practical deployment in traffic monitoring and law enforcement applications, and supporting automated Intelligent Transportation Systems (ITS).

AI is revolutionizing fashion by automating clothing categorization and trend identification. This project classifies outfits into casual, sports, and formal categories using TensorFlow, ensuring accurate classification while identifying key attributes like gender, color, and outfit type. Convolutional Neural Networks enhance pattern recognition, improving classification accuracy, refining fashion suggestions, facilitating product filtering on e-commerce platforms, and providing tailored outfit recommendations that lead to an improved shopping experience and higher customer satisfaction.