Authors - Hiruni Samarage, Pumudu A. Fernando Abstract - Tertiary education institutes increasingly face challenges in managing high volumes of student inquiries related to admissions, courses, fees, and scholarships. Traditional inquiry-handling mechanisms and rule-based chatbots often struggle with scalability, delayed responses, and limited understanding of complex or unstructured queries. While recent advances in large language models (LLMs) offer promising opportunities, many existing academic chatbot implementations continue to lack semantic retrieval, session continuity, and personalization. This paper presents the design, implementation, and evaluation of an AI-based semantic chatbot prototype tailored for tertiary education environments. The prototype integrates retrieval-augmented generation with a large language model to enable context-aware responses across multiple institutional knowledge domains through semantic retrieval of structured knowledge representations. The system was evaluated using accuracy, precision, recall, robustness to query length variations, and retrieval effectiveness metrics. Experimental results demonstrate an overall accuracy of 86%, with precision and recall values of 89% and 91%, respectively. Robustness testing shows consistent performance across paraphrased and variable-length queries, while response times remained within acceptable limits for real-time academic support. User testing further indicated positive usability and response relevance outcomes. These results confirm the feasibility and effectiveness of applying semantic retrieval and LLM-based reasoning to scalable inquiry management in tertiary education contexts.
Authors - Ndaula Kelvin, Wu Jun Abstract - Multimodal sentiment analysis often fails due to the modality gap between semantic text images and GIFs. To address these challenges this paper introduces Fusion Core a novel hardware agnostic heterogeneous pipeline that bridges the gap between high level AI with low level systems engineering to facilitate the deciphering of combined sentiment of these three modalities. Through the integration of GPGPU accelerated OpenCL kernels for 3D temporal extraction with an ONNX/DirectML inference engine which ensures cross platform portability. To address the issue of inconsistent real world data distributions the preprocessing system was introduced with an adaptive multi-head attention mechanism for late feature fusion. The ablation studies performed also revealed the integration of spatiotemporal GIF layers resolves contextual ambiguities missed by static analysis (Text, Images). The extensive testing on a balanced dataset of 13,964 samples the model achieved a 100% success rate showing the robustness of the proposed model for industrial scale deployment.
Authors - Tumiso THULARE, Keneilwe Jeannette MAREMI Abstract - This paper presents findings from a pilot e-Participation implementation conducted in partnership with the Mpumalanga Department of Cooperative Governance, Human Settlements and Traditional Affairs (CoGHSTA) and the City of Mbombela over two years. The pilot project focused on enhancing the municipality's capacity to implement and sustain e-Participation initiatives. It also assessed the current state of public participation and explored the opportunities and challenges of adopting digital participation mechanisms in South African local government. A qualitative research approach was used, involving a scoping review and engagement sessions with municipal officials from various units, including public participation, communications, ICT, policy, and governance. The scoping review identified theoretical challenges to e-Participation in South African municipalities, while engagement sessions examined institutional experiences, governance processes, and the City of Mbombela's readiness for digital participation. The findings revealed that the municipality shows policy alignment and has partially adopted digital participation tools such as social media, municipal websites, and mobile communication channels. However, e-Participation implementation faces challenges such as the digital divide, limited ICT infrastructure, low digital literacy, institutional capacity constraints, poor coordination, and insufficient funding. The study further found that public participation still relies heavily on traditional methods, with digital platforms mainly used for sharing information instead of fostering citizen empowerment or collaborative governance. The paper concludes that while e-Participation can enhance transparency, accountability, and inclusive governance in South African municipalities, its success requires integrated technological, institutional, and participatory reforms. The findings provide practical guidance for municipalities aiming to enhance digital public participation in resource-constrained settings.
Authors - Shola Usharani, Gayathri Rajakumaran, Braveen manimozhi, Anjana Devi Nandam, Kindinti Karthik, Prabhakaran mohan Abstract - The closed-loop anesthesia delivery (CLAD) development is a significant advancement in modern anesthesiology, offering automatic control of anesthetic administration to maintain optimal patient states during surgical procedures. The article analyses the limitations of existing systems—such as sensor errors, limited adaptability, and system unreliability—by integrating through IoT based control high-accuracy EEG monitoring system with a robust and novel closed-loop control algorithms. From this system the real time EEG signals are continually monitored, analyzed, filtered, and converted into a BIS spectral Index (BIS) values. The target BIS is continuously updated by fuzzy logic, which modulates the drug delivery to maintain sedation levels between 40 and 60. A PID controller used for the BIS error calculation to determine the precise anesthetic levels to be delivered. This amount level is then converted into drug level concentrations to drive the syringe pump connected to the IoT integrated system using stepper motor to administer the drug to the patient. The serial monitor displays all information in real time, including BIS values, PID output, calculated dosage in mg/sec and ml/sec, and the number of motor pulses required for the infusion. This prototype demonstrates enhanced precision and safety over manual control, offering a scalable and cost-effective solution for automated anesthesia delivery, thus avoiding the limitations of existing model.
Authors - Norbert Annus Abstract - This study presents a secondary learning analytics analysis of student log data generated by the Learn with M.E. educational software. While previous evaluations of the system focused mainly on effectiveness, diagnostic accuracy and user feedback, the present paper examines behavioural indicators recorded during arithmetic practice. In this study the analysis focused on calculation time, answer correctness, difficulty level, first-try success, "Preview" use and student-level behavioural profiles. The results showed that incorrect answers were associated with substantially longer calculation times than correct answers. Higher difficulty levels generally showed lower correctness rates and longer median calculation times. First-try attempts were also strongly related to successful task completion. "Preview" use was relatively rare, but it was associated with longer calculation time, higher average difficulty level and lower first-try correct rates, suggesting that it can be interpreted as a help-seeking indicator. The student-level aggregation identified four behavioural profiles: fast trial-and-error learners, help-seeking learners, mixed-profile learners and support-needed learners. The findings indicate that Learn with M.E. log data can be transformed into interpretable learning analytics indicators and behavioural patterns that support teacher decision-making and personalised mathematics instruction.
Authors - Sheromiga Anandajothy, Aathipan Murugaverl, Harinda Fernando, Sarangan Rukminikanthan, Abishathan Thayaparan, Tharaniyawarma Kumaralingam Abstract - Modern malware increasingly employs packing, encryption, polymorphism, staged payload delivery, modular execution, and behavioural evasion techniques to bypass traditional signature-based security systems. While static analysis enables rapid inspection of suspicious binaries, it often performs poorly against heavily obfuscated samples. Conversely, dynamic analysis provides rich runtime evidence but introduces computational overhead, operational latency, and anti-sandbox challenges. This paper presents Mutated Malware Protector, a lightweight hybrid framework for detecting obfuscated and modular malware through static Portable Executable (PE) feature analysis, obfuscation-aware scoring, behavioural risk approximation, modular stage inference, and explainable artificial intelligence (XAI). The framework is designed as a practical analyst-facing pipeline rather than a single classifier. It integrates engineered PE features, anomaly scoring using Isolation Forest, entropy-based obfuscation indicators, a LightGBM static classifier, behavioural approximation, and a modular correlation component that estimates staged roles such as dropper, loader, and payload. Experimental evaluation shows that the prototype achieved 89.00% accuracy, 89.28% precision, 88.81% recall, and 89.04% F1 score. The proposed architecture offers a scalable path toward future integration with full sandbox telemetry and enterprise malware response workflows.
