ABSTRACT. The topic of this doctoral dissertation stems from the fact that, despite previous initiatives and projects aimed at digitalizing education, the systematic application of artificial intelligence in the educational process remains neither clearly defined nor sufficiently explored within the context of the Croatian education system.

The scientific contribution of this dissertation will be the development of a new, efficient model for applying artificial intelligence in the educational process, based on empirically validated results and tailored to the specific characteristics of the Croatian educational context.

The research aims to address the existing gap in the literature and provide a solid scientific and professional foundation for the further implementation of AI in the education system. The findings will contribute both to the theoretical understanding of the impact of AI technologies on educational processes and to practical recommendations that will support schools in achieving an effective intelligent transformation.

ABSTRACT. The research focuses on the application of multimodal deep learning models for cryptocurrency price prediction. A multimodal model involves the integration of market data (such as closing prices within specific time intervals and trading volume) with on-chain blockchain metrics (such as the number of active addresses and daily transaction count). The model is based on LSTM neural networks trained to process market and on-chain data in parallel. The findings of this study may serve as a foundation for the development of more advanced analytical systems in the field of financial technologies. Additionally, the work contributes to the existing literature by applying the LSTM architecture to structured blockchain data—an area that remains relatively underexplored.

ABSTRACT. The doctoral research aims to examine how the adoption of artificial intelligence (AI) technologies contributes to the digital transformation of small and medium-sized IT enterprises in Bosnia and Herzegovina. The focus is on understanding how AI enhances productivity and business process efficiency, while also identifying key challenges in implementation, such as limited financial and human resources. The study will employ a quantitative research approach, combining primary data collected through structured online surveys targeting IT company managers with secondary data from literature, legal frameworks, consultancy reports (e.g., Deloitte, Eurostat, KPMG), and statistical databases. Data analysis will include descriptive statistics, t-tests, ANOVA, correlation and regression analyses. In addition, qualitative insights will be supported through SWOT analysis and thematic interpretation. This research is expected to contribute to both theoretical and practical understanding of AI’s role in digital transformation. It will offer concrete recommendations for improving AI adoption in the IT sector and serve as a foundation for future policy development aimed at fostering innovation, competitiveness, and digital readiness in Bosnia and Herzegovina. Recommendations for further research will address long-term workforce impacts, comparative international analyses, and implications for data security and privacy.

ABSTRACT. The increasing integration of information technology into all industries is changing machines, processes and affecting workplace safety. The share of electronic components and software in production systems is constantly growing. Progress in the global connectivity of industrial systems brings with it the challenge of collecting, storing and analysing data. Increased digital connectivity requires continuous investment in the latest tools, hardware and innovative strategies to protect networks from cyber attacks. The topic of this lecture is safety in research, testing, implementation and application of Industry 4.0 technologies.

ABSTRACT. Concept mapping is a powerful method for visually organizing and representing knowledge. This presentation discusses how concept maps can be automatically extracted from semantic hypergraphs using a series of simple rules. A semantic hypergraph’s recursive structure allows for the representation of both concepts and their relationships, which can then be converted into triples (subject, predicate, object) to form a concept map.

It is demonstrated how semantic hypergraphs, built from text using basic NLP techniques like dependency parsing, semantic role labeling, and coreference resolution, can be transformed into concept maps by identifying key concepts and their relationships. The process of transforming a semantic hypergraph into a concept map involves three main steps: identification, classification, and aggregation. During identification, candidate semantic hyperedges are selected; in classification, the subject and object of triples are determined; and in aggregation, hyperedge components are combined to generate a unified triple. The final step involves visualizing the concept map, where triples are depicted as nodes connected by links.

It is shown through examples how this process works, focusing on transforming sentences into concept maps at various levels of detail. Additionally, concept maps generated by the system are compared with human-generated maps, and the effectiveness of simplification rules for reducing complexity is analyzed.

ABSTRACT. Medical image analysis plays a crucial role in the interpretation and processing of medical images to support diagnosis, treatment, and research. Continuous advancements in machine learning have significantly improved the accuracy and efficiency of this field. Despite substantial improvements, numerous challenges still remain. This presentation will provide insight into the latest achievements in medical image analysis, with a focus on machine learning methods, while also addressing existing challenges and future research directions.

ABSTRACT. This article describes a workshop designed to introduce artificial intelligence (AI) concepts to elementary school students using interactive tools and discussions. The workshop, attended by over 50 fifth and sixth-grade students in several sessions, aimed to demystify AI principles by incorporating the "Brain in a Jar" concept and exploring the "Chinese Room Argument". The workshop's centrepiece was a "smart box", made especially for the workshop, symbolizing AI, which engaged students in understanding how machines learn and make decisions. Through hands-on activities, such as interaction with NAO robots and the "smart box", students learned about AI's basic concepts like object recognition and decision-making processes. The "Brain in a Jar" concept was used to provoke discussions on the nature of intelligence and the potential for machines to replicate human-like cognitive functions. Meanwhile, the "Chinese Room Argument" raised ethical and philosophical questions about AI's understanding versus true comprehension. Pre-workshop and post-workshop surveys were conducted to gauge participants' understanding and perceptions of AI. Initial findings indicate increased awareness and curiosity about AI technology among students, highlighting the effectiveness of interactive learning experiences in fostering technological literacy and critical thinking. This article provides insights into effective pedagogical approaches for integrating AI concepts into elementary school curricula, highlighting interactive tools and philosophical discussions. The presented model offers teachers innovative ways to introduce complex technological concepts, improve digital literacy, and incorporate new technologies into education.

ABSTRACT. This paper compares two approaches to robotics education that are character-ized by a strong practical orientation and a focus on theoretical knowledge transfer. The aim of the study is to analyze the effects of these different teaching methods on students' skills and to shed light on the associated challenges. In addition, the rapid development of robotics and increasing digitalization will be analysed as factors that place new demands on academic edu-cation. The study examines the integration of robotics into the curricula of institutions of higher education, contrasting and comparing the pedagogical methods of the University of Applied Sciences, where industry-specific, hands-on training is provided, and the University of Mostar, where robotics is included in more general computer science and engineering curricula. The paper establishes the strengths and weaknesses of both approaches, namely in curriculum development, lab facilities, industry partnerships, and projects. Technological progress is taking place at a rapid pace, meaning that curricula need to be continuously adapted, innovative teaching methods developed and teaching materials regularly updated. In the field of robotics, this poses significant challenges for both universities of applied sciences with their practice-oriented education and universities with their strong scientific focus, as both teaching staff and the entire university organization are under increasing pressure. The study also addresses the ethical dimensions of robotics and AI, advocating ethical standards in robotics studies to foster so-cially responsible innovation. The paper also recommends a mixed method-ology that combines industry engagement with AI-based research, exposing students to practical experience and forward-thinking innovation.

ABSTRACT. Humanoid robots in education present both promising opportunities and considerable risks. This review explores the use of humanoid robots in formal education, focusing on pedagogical approaches, impact, challenges, and ethical considerations. Most studies employed constructivist and socio-cultural approaches, such as problem-based and game-based learning. The results revealed that humanoid robots enhance learning effectiveness, student engagement, motivation, enjoyment, confidence, and 21st-century skills. However, privacy, data security, and reduced engagement with people due to over-reliance on robots remain ethical concerns. Implementation challenges include technical issues, high costs, and a lack of trained personnel. The review offers recommendations for implementation, future directions, and research to guide the responsible and effective integration of robots into formal education.

ABSTRACT. Digital literacy is considered the fourth analytical skill that students today should possess, and many initiatives promote its development in children from an early age. It is important to emphasize that the focus should be on using technology and understanding the basic concepts of computational thinking (CT). There is a lack of research on teaching CT in kindergarten, partially due to the absence of age-appropriate technology interfaces that would enable children to engage in more complex projects. On the other hand, CS unplugged teaching methods for fundamental computer science concepts have already been proven successful, even in kindergarten. This paper describes research designed based on Papert's experiments, focusing on using the "turtle" as a real-world object in teaching sequence, branching, and loop algorithms. The research was conducted in three kindergarten groups involving 34 children aged 3 to 6.

ABSTRACT. The rise of Artificial Intelligence (AI) and Large Language Models (LLMs) has created new opportunities for intelligent educational systems, particularly in customizing and personalizing learning content. This paper explores strategies, methodologies, and lessons learned in developing adaptive educational content using LLMs within Moodle, a widely used Learning Management System (LMS). Drawing on CARNET’s large-scale educational transformation initiatives, the paper demonstrates how LLMs dynamically adapt learning materials to meet specific user needs, knowledge gaps, and learning styles. By leveraging AI-driven content that responds to student performance during assessments, LLMs generate tailored learning materials and personalized pathways. The process includes training and fine-tuning LLMs on domain-specific knowledge, developing content templates, and integrating them into the LMS. This enables differentiated learning experiences unique to each user’s progress. Key challenges, including technical complexities, human factors, and cross-stakeholder collaboration, are also discussed. Practical insights are provided on content relevance, system integration, user adoption, and the role of intelligent tutoring systems (ITS) in supporting adaptive learning at scale. This paper offers a roadmap for institutions to incorporate LLMs and AI-driven customization into their learning ecosystems.

ABSTRACT. In this paper, we investigated the factors affecting human accuracy in geolocalizing wildfires based on visual imagery. Through an online survey with 71 participants of diverse demographics, we evaluated three key hypotheses regarding wildfire location estimation. Participants were presented with images captured by the FireDetectAI system during summer 2022 and asked to mark estimated fire locations on a map using a custom Leaflet-based interface. Our analysis revealed that incorrectly marked fire alarms significantly reduced location accuracy (mean distance error of 7.86 km versus 9.89 km for other scenes). Camera misalignment also negatively impacted geolocalization precision, with misaligned scenes showing higher error values when outliers were excluded. The presence of landmarks improved estimation accuracy (mean distance error of 7.84 km versus 10.37 km for scenes without landmarks). Statistical analysis of the results reveals inherent challenges in relying solely on human visual assessment for wildfire location estimation, particularly in featureless terrain, and suggests that improved annotation accuracy, proper camera alignment, and landmark presence could enhance human geolocalization performance. These insights have important implications for the design of wildfire monitoring systems and emphasize the potential benefits of supplementing human assessment with automated detection and geolocalization technologies for more effective early warning and response.

ABSTRACT. Abstract. The development of large language models, such as Chat GPT, has significantly advanced in machine translation (MT), partic ularly in multilingual text generation and interpretation. However, as sessing the translation quality of these models across diverse languages and domains remains a complex challenge. This study proposes a com prehensive approach to assessing the translation performance of Chat GPT4o, designed to capture the nuances of its multilingual capabilities. The methodology combines traditional evaluation metrics (e.g., BLEU, METEOR, TER, ChrF, COMET, and BLEURT) with human-centered evaluation methods that prioritize linguistic quality and contextual accu racy. The evaluation spans four languages (English, Italian, French, and Croatian) and covers publicly available domain-specific texts in religion, medicine, economics, and law, comprising a total of 676 sentences. The human evaluation was conducted by two experts specializing in English French and English-Italian language pairs, as well as 26 English lan guage undergraduates. The assessment was based on several key criteria, including linguistic nuance, grammatical and syntactic precision, termi nological consistency, contextual appropriateness, textual coherence and flow, and adherence to punctuation and formatting norms. The research addressed the following questions: Does ChatGPT provide equally accu rate translations across all languages? Are there significant differences in translation quality depending on the domain? What are the most com mon discrepancies or errors when compared to reference translations? What is the relationship between automatic metrics and human judg ments? Could ChatGPT potentially outperform a human translator in certain contexts? The analysis highlights the model’s strengths and limi tations in various linguistic settings and identifies areas for improvement to enhance translation reliability. This study contributes to continuous ef forts to improve the understanding and evaluation of machine translation systems, offering a refined perspective on assessing AI-driven translation tool in multilingual contexts.

ABSTRACT. Advanced intelligent systems, particularly Large Language Models (LLMs), are significantly reshaping financial practices through advancements in Natural Language Processing (NLP). However, the extent to which these models effectively capture and apply domain-specific financial knowledge remains uncertain. Addressing a critical gap in the expansive Indian financial context, this paper introduces CA-Ben, a Chartered Accountancy benchmark specifically designed to evaluate the financial, legal, and quantitative reasoning capabilities of LLMs. CA-Ben comprises structured question-answer datasets derived from the rigorous examinations conducted by the Institute of Chartered Accountants of India (ICAI), spanning foundational, intermediate, and advanced CA curriculum stages. Six prominent LLMs—GPT-4o, LLAMA 3.3 70B, LLAMA 3.1 405B, MISTRAL Large, Claude 3.5 Sonnet, and Microsoft Phi 4—were evaluated using standardized protocols. Results indicate variations in performance, with Claude 3.5 Sonnet and GPT-4o outperforming others, especially in conceptual and legal reasoning. Notable challenges emerged in numerical computations and legal interpretations. The findings emphasize the strengths and limitations of current LLMs, suggesting future improvements through hybrid reasoning and retrieval-augmented generation methods, particularly for quantitative analysis and accurate legal interpretation.

ABSTRACT. The integration of Extended Reality (XR) technologies, including Virtual Reality (VR) and Augmented Reality (AR), has emerged as a transformative approach in medical education, surgical planning, and diagnostics. This mini-review consolidates advances in XR applications within medical training and planning, highlighting innovations in prenatal ultrasound, cerebrovascular surgery, and synthetic medical image generation. By analyzing current methodologies, we explore the potential of XR to simulate rare clinical conditions, improve diagnostic accuracy, and enhance surgical safety. The review also addresses technical challenges, such as image realism and system scalability, and proposes pathways for future developments in XR-driven personalized medicine.

ABSTRACT. The development of artificial intelligence (AI) is bringing sig- nificant changes to music education, particularly through a personalized learning approach at the primary level. This paper explores the applica- tion of AI technologies in adapting teaching content to the cognitive and affective characteristics of students. AI enables flexibility in learning, ad- justs the pace, and provides systematic real-time feedback. By using AI tools, students have the opportunity for interactive and adaptive learn- ing, while teachers can more effectively monitor their progress through specific case studies that demonstrate the pedagogical impact of AI in primary school music education. The benefits include increased motivation, the development of creativity, and more precise evaluation of musical competencies. Despite numerous advantages, challenges such as ethical concerns, the potential depersonal- ization of teaching, and the need for teacher training remain key aspects of AI implementation in education. The research findings confirm that AI can significantly enhance music education if implemented in accordance with pedagogical principles and ethical guidelines. This topic is significant as it explores how AI can tailor teaching to the individual needs of students, which is one of the key trends in modern education. Additionally, it allows for the exploration of specific AI tools that can assist teachers in educating children.

ABSTRACT. The Covid-19 pandemic caused a situation in which both educators and students had to adapt as soon as possible, and a new educational context known as distance learning has become a daily routine for teaching all courses. However, some courses were easily adapted to the new type of learning, while for some courses it was very difficult to provide an efficient online teaching concept. This was especially true for engineering courses such as digital systems design, which requires working in laboratories and performing live demonstrations on digital systems and Field Programmable Gate Arrays (FPGA) boards. One of the biggest problems was that the students did not have access to the boards as they were physically located in the labs. It is not easy for students to buy the equipment for their learning purposes, as the boards are specialized and costly. The idea of this paper is to provide an overview of how researchers in the current state-of-the-art have approached this problem and how they applied distance learning methods to implement systems for students to access remote FPGA boards. Several methods for remote lab implementations were found and described and contrasted by comparing their advantages and disadvantages. Finally, a conclusion is drawn, taking into account potential directions that could be further explored as future work, not only as a tool for practical emergency situations, but also as an additional for in-person teaching.

ABSTRACT. The rapid development of market demand and the pursuit of sustainable industrial practices require a change in production and business models. This transfor-mation is being driven by new technologies such as cyber-physical systems (CPS), the Internet of Things (IoT) and cloud computing. Human-centric digital twins (HCDTs) play a crucial role in this transformation as they put people at the heart of digitalization. By integrating human data, ergonomics and real-time inter-actions, HCDTs not only optimize efficiency, but also improve workplace safety, employee well-being and adaptability. This approach supports the development of flexible, customized and resilient production systems that align with the goals of Industry 4.0 and sustainability. The learning factory environment at the Universi-ty of Split, Faculty of Electrical Engineering, Mechanical Engineering, and Naval Architecture (FESB), provides an applied education platform that combines aca-demic knowledge with industrial requirements. This program offers students, and industry partners the opportunity to engage in real-world projects focused on the implementation of HCDT. Participants gain hands-on experience using sensors and CPS to collect and analyze data and use modeling and simulation tools to cre-ate digital twins. By working in a collaborative environment, students develop critical problem-solving skills while industry partners explore innovative solu-tions to optimize production processes. This hands-on training provides the par-ticipants with the knowledge and technical skills they need to integrate new tech-nologies into engineering and industrial applications, driving innovation and competitiveness in the evolving manufacturing landscape.

ABSTRACT. Autism Spectrum Disorder (ASD) is a broad neurodevelopmental condition affecting many young learners. Educational efforts are made to meet their needs, overcome deficits and increase their well-being in classroom and everyday activities. The aim of this paper is to review the existing literature regarding the impact of 3D-printing activities and 3D-printed educational models on the learning, engagement, and social integration of children and adolescents with autism. Over 100 relevant articles from Google Scholar and Science Direct are reviewed, with 34 selected for an in-depth analysis. Among them, 12 fully met the study criteria. Best practices in this field are identified in order to analyze challenges and propose solutions for optimizing the application of 3D printing technology in special education. The features and findings of those investigations are illustrated in two tables, including factors such as the type of assistive technology and software, research methodology and target groups, lesson design specifics, observed outcomes, and social needs addressed. Protocols for the design of special education lessons are also reviewed and discussed.

ABSTRACT. Artificial intelligence in education is bringing significant changes to the way learning and teaching is done, and its application in mathematics teaching opens up numerous opportunities for improving the educational process. There is a clear connection between mathematics and artificial intelligence, as many AI algorithms and technologies base their functionalities on mathematical principles, such as statistics, linear algebra, and optimization theory. This article explores the mathematical foundations that support the development of artificial intelligence, emphasizing how mathematics provides the foundation for understanding and implementing AI technologies. The emphasis of the article will be on the application of artificial intelligence in mathematics teaching, with the aim of improving the experience of students and teachers. AI enables personalized learning, where mathematical models are used to adapt tasks and teaching content to the specific needs of each student. This research will explain how artificial intelligence can help optimize teaching methods, improve learning efficiency, and support the development of metacognitive skills in students. Although the advantages of applying artificial intelligence in education are significant, it is important to consider the challenges and obstacles that may be encountered in implementing this technology in order to fully utilize its potential in mathematics teaching.

ABSTRACT. This paper presents an exploratory data analysis and prediction model for fuel prices in the Federation of Bosnia and Herzegovina (FBiH) relative to oil price movements. The study focuses on the Super 98 fuel type and utilizes data collected from the Federal Ministry of Trade of BiH's eOPC application from December 2018 to September 2023. The analysis employs various data visualization techniques and statistical methods to identify trends and correlations. A recurrent neural network, specifically a Long Short-Term Memory (LSTM) model, is constructed to predict future fuel prices based on historical data and the movements of BRENT and WTI oil prices. The results indicate a high correlation between fuel prices in FBiH and BRENT oil prices, with the LSTM model demonstrating high predictive accuracy. This research provides valuable insights for decision-makers in economics and energy sectors and suggests potential future research directions, but also publicly publishes and analyzes a official Ministry dataset over a period of 5 years for 716 gas stations and six fuel types.

ABSTRACT. With an emphasis on targeted advertising and public communication this study researches the effects of AI tools on communication strategies of or-ganizations. While testing Hypothesis posits that AI enhances targeted ad-vertising through successful data analysis. The principal hypothesis posits that AI technologies augment communication strategies through personal-ized, timely, and data-driven public engagement. Using a quantitative ap-proach with 51 participants, the study employed descriptive statistics, Kruskal-Wallis tests, and regression analysis. Results show that 73.7% of the variance in AI tool effectiveness is explained by budget readiness, in-ternal communication frequency, and external communication quality. Both hypotheses were confirmed, demonstrating AI's critical role in improving communication strategies and targeted advertising outcomes within organizations.

ABSTRACT. Artificial intelligence (AI) is not an entirely new phenomenon, but in recent years, it has experienced rapid expansion in all spheres of life, including business. AI application improves business but also cause significant challenges to providing a transparent, safe, and ethical AI application. The development of AI tools can help insurers improve underwriting and monitor and predict risks better. Using digital tools has simplified the claims process, and AI applications can help detect insurance fraud. With AI applications, insurers speed up their

operational processes, which leads to cost savings and improved customer expe- rience. In addition to these benefits, the authors analyze some of the legal chal- lenges AI application can cause, like the lack of regulatory compliance and due

diligence, ethical matters, bias and discrimination, human rights violations such

as data protection and privacy, etc. Furthermore, this paper provides a brief over- view of the regulatory framework for the application of AI in the European Un- ion. To provide a comparative overview with the EU insurance market, the au- thors conducted a statistical analysis of the application of AI and digital tools in

the insurance market in Bosnia and Herzegovina. Based on the analyses, the au- thors conclude that the future of AI in insurance depends on the optimal ratio

between business benefits and transparent and ethical legal regulation. Regard- less of the legal challenges, implementing AI business models is a matter of com- petitive advantage in the insurance market.

ABSTRACT. The research demonstrates an application of a no-code AI model in the field of international MTPL insurance claims. The model uses natural language processing (NLP) to analyze a custom-made case law database, create summaries, and provide translations into the Croatian language when needed. The AI model was developed using a Generative pre-trained transformer (GPT) and the effectiveness of the model was evaluated through two studies in which the participants were given a hypothetical cross-border insurance claim. The pilot study revealed, and the follow-up study confirmed, that the GPT significantly improves research efficiency and helps overcome language barriers in international MTPL claims handling.

ABSTRACT. This study explores the application of the Random Forest algorithm for classifying dominant team roles based on Belbin’s Team Roles Self-Perception Inventory. A synthetically generated dataset of 100,000 instances was used to train and evaluate the model, ensuring that the distribution of team roles reflects real workplace environments. The results demonstrated high classification accuracy, with an overall precision of 0.98725. The study found that roles with distinct behavioral characteristics, such as shaper and coordinator, were easier to classify, whereas roles with overlapping traits, such as monitor evaluator and plant, posed greater challenges.

The findings have significant implications for modern workforce management. By utilizing machine learning for team role classification, organizations can optimize team formation, enhance collaboration, and improve productivity. The model offers a data-driven approach to aligning employees with roles that match their strengths, reducing subjective decision-making in human resource processes.

Despite its effectiveness, the study acknowledges limitations, such as the reliance on synthetic data and the need for further validation using real-world datasets. Future research should explore deep learning models and integrate natural language processing techniques for improved role prediction. This study highlights the potential of AI-driven HR analytics to create balanced, high-performing teams in contemporary business environments.

ABSTRACT. „DIGIPHY“ is the research project aiming to realize immersive and intuitive XR communication and interaction in the visual representation of dynamically updated digital twin (DT), spatially and temporally synchronized with a physical smart space (equipped with sensors and actuators). Through the research process, different case studies are covered thus showing the evolution of use cases within different vertical domains.