Conference Agenda

Automated grading systems improve efficiency in
evaluating exams, but handwritten code grading remains timeconsuming
and subjective. In this paper, we introduce an automated
grading system for handwritten programming code using
a TESSERACT-based OCR model, achieving 90% accuracy
on a real-world dataset of 500 handwritten code samples. To
ensure practical relevance, we introduce the AUI-HandCode
dataset, a novel collection of handwritten Python code gathered
from Introduction to Computers course exams administered
between 2009–2013 at the School of Science and Engineering, Al
Akhawayn University (AUI), in Ifrane, Morocco. These pen-andpaper
exams, conducted under 60–90 minute timed conditions,
reflect authentic student challenges in solving algorithmic problems
without digital assistance.
Our system provides faster, fairer, and more objective evaluations,
reducing human bias and inconsistencies. Ground-truth
labels were validated by an instructor following strict annotation
guidelines, with ambiguous cases resolved using a syntax-aware
checker. Unlike prior work limited to synthetic or controlled
handwriting, [1] our model addresses real-world challenges such
as inconsistent indentation, cursive symbols, and varying handwriting
quality. While performance on synthetic benchmarks
reached 98% accuracy, real-world accuracy of 90% highlights
meaningful trade-offs involved in practical deployment.
The proposed system supports handwritten coding assessments,
which can reduce academic dishonesty and promote
authentic problem-solving. Our findings demonstrate the model’s
potential for scalable use in educational settings, enabling
instructors to maintain academic integrity while significantly
reducing grading workload. Limitations related to extreme
handwriting variability will be addressed in future work through
adversarial training and expanded data collection.

E-learning systems in bandwidth-constrained environments face persistent challenges in traffic prioritisation, quality of service (QoS), and security. Conventional best-effort networks cannot reliably support the latency-sensitive traffic generated by modern digital learning platforms, particularly in developing countries with limited infrastructure. This paper presents a systematic literature review following the PRISMA 2020 methodology, analysing 26 peer-reviewed studies published between 2018 and 2025. Drawing on the identified gaps, a three-layer adaptive SDN-based framework is proposed, integrating application-aware traffic classification, dynamic bandwidth allocation, and security management. The framework is validated through Mininet-based network emulation across three load scenarios. Results show jitter reductions of up to 100% for video lecture traffic, a 74.9% RTT improvement for collaborative tools under overload conditions, and restoration of 98.7% video throughput within three seconds of a simulated UDP flood attack. The framework operates over existing IP infrastructure, making it suitable for resource-constrained educational deployments.

AI technologies have disrupted educational practices and workforce development across sectors, including significant impacts in higher education. Like past disruptions resulting from rapid technological change, the use of AI in higher education has not only changed the means by which learning and work are accomplished but has also raised challenging questions about educators’ personal and professional identities. In particular, the conceptualization of AI as a collaborator or amplifier of human activity can be regarded as either an opportunity for more meaningful and rewarding work or, alternatively, as a dehumanizing tool for marginalizing or replacing human workers altogether. Such considerations have resulted in calls for the development and promotion of human-centered AI literacy frameworks and educational programs that better empower humans to participate in AI development and adoption as critically reflective and ethical users of emerging technologies. The capability approach is a framework for human development that has been applied extensively to topics relevant to AI literacy, including education and technology. This paper describes a generalized model of AI literacy pedagogy that extends considerations of human agency and well-being, as understood within a capability approach, to AI education and training in higher education.

Large Language Models (LLMs) are becoming a common part of higher education, but many students still find it difficult to turn broad AI ethics principles into practical decisions for real projects. We present EthicsAdvisor, a classroom tool designed to support this process through scenario-based ethical guidance. The system combines a fine-tuned Llama 2 model with a lightweight FastAPI + Streamlit interface and produces structured outputs covering key issues, stakeholders, trade-offs, possible mitigations, and short explanations. We tested the tool in an exploratory pilot study with 20 data-science students at the University of Koblenz. Using Likert-scale questions and open-ended responses, we examined how students perceived its usefulness, clarity, relevance, and effect on their confidence after hands-on use. Overall, students responded positively. Many described the tool as useful, practical for project work, and helpful in drawing attention to risks they had not fully considered before. Qualitative feedback also showed that students appreciated the structured format and the simple workflow, while first-call latency and occasional verbosity were seen as the main drawbacks. Although this study does not demonstrate learning gains or causal effects, it provides early evidence that structured LLM-based support can be a practical and well-received way to encourage ethical reflection in project-based data-science education. We conclude with design implications and directions for future work, including objective learning measures, baseline comparisons, expert review of output quality, and deployment across multiple courses.