JITA - APEIRON

Implementation of a Meshtastic Gateway System With a Local Database for Iot Applications

2026-01-05T14:39:16+01:00

This paper presents a system for reliable collection, filtering, and processing of data from a LoRa Meshtastic decentralized network, developed for use in remote areas with weak or no mobile network coverage. The core idea stems from the need to enable efficient exchange of small data packets at intervals, without relying on expensive and often unavailable internet infrastructure. The key innovation lies in the implementation of the Meshtastic gateway concept, which provides internet access via HTTP requests, while the developed database model ensures continuity and reliability of data transmission. Data arriving in the network as unstructured messages are extracted using regular expressions, transformed into JSON format, and sent to the visualization platform Grafana, while simultaneously being stored in a local database for later queries, research, and analysis. The system’s reliability is further enhanced by introducing a two-layer acknowledgment mechanism (Meshtastic ACK_APP and remote API application-level ACK), as well as an offline mode that logs undelivered messages and their causes through flags and error records. This ensures resilience to data loss and enables seamless operation continuation after connection interruptions.

CNN-Based Road Sign Recognition for Driver Assistance

2026-01-05T14:49:21+01:00

Considering established relevance to the GTSRB dataset, it is important to emphasize that research investigates the effectiveness of convolutional neural networks (CNN) in the field of road sign recognition. Following that wide range of techniques for comprehensive preprocessing pipelines were implemented, including data normalization and augmentation as well as resizing images. The CNN model has demonstrated the ability to overcome adverse conditions across multiple road sign classes, demonstrating outstanding scores against the performance metrics used in testing and evaluation process. Model achieved classification accuracies exceeding 99% across most categories. Nevertheless, in certain classes there is presence of performance metric decline related to the inaccurate visualization and contradiction of features. The crucial role of the preprocessing phase has been highlighted while the implementation of the CNN model has been identified as one of the most reliable approaches in the field of road sign recognition. However future implications must be considered to achieve the full potential of the model. Some of the crucial contributions for the future will be introducing real life variation in the dataset. On the other hand, occlusion, lighting and weather conditions are the important factors that should be brought into focus.

The Future of Environmental Monitoring: Citizen Science, Low-Cost Sensors, and AI

2026-01-05T15:22:30+01:00

The increasing availability of low-cost environmental sensors and the integration of Artificial Intelligence (AI) into data processing are reshaping citizen-driven environmental monitoring. This study explores public engagement with such technologies, focusing on the willingness of different population groups to participate in monitoring activities and the trust they place in AI-supported sensor data. By combining citizen science approaches with AI-assisted interpretation, the research aims to assess how individuals perceive the reliability, usefulness, and accessibility of environmental information. A quantitative survey was conducted using a 15-item online questionnaire distributed to four groups: university students, general citizens, active participants in citizen-science projects, and IT/data professionals. The survey included multiple-choice, Likert-scale, and short open-ended questions to capture a comprehensive picture of familiarity with environmental monitoring, attitudes toward participation, and perceived role of AI in enhancing data credibility. The collected data were analyzed using descriptive statistics and comparative group analysis. All anonymized data, survey instruments, and analysis files have been made publicly available in the AIMIS-Survey-2025 GitHub repository (https://github.com/oljak-cyber/AIMIS-Survey-2025), ensuring reproducibility and transparency. Results indicate that participants are generally willing to engage in citizen-led monitoring, with IT and active citizen-science participants demonstrating the highest levels of trust and readiness. AI-assisted validation of sensor data was perceived as a significant factor in enhancing confidence and interpretability, particularly among technically proficient respondents. Main barriers identified included cost, lack of knowledge, and time constraints, highlighting the importance of accessible technology and educational guidance for broader adoption. Overall, the study underscores the potential of combining low-cost sensors with AI tools to empower citizens, improve environmental awareness, and generate reliable datasets for informed decision-making. Future initiatives should focus on public education, transparent AI models, and scalable sensor deployments to maximize engagement and ensure data quality.

The Role of AI Assistants in Supporting Teachers

2026-01-06T10:32:49+01:00

The topic of this research is the role of AI assistants in supporting teachers within contemporary education. The primary aim is to examine how teachers perceive the potential of AI-based tools, which specific tools they use, in which instructional contexts they apply them, what challenges they recognize, and which ethical concerns they consider crucial for their safe and responsible integration into the educational process. The study was conducted through a survey among primary and secondary school teachers, followed by a combination of quantitative and qualitative data analysis. The findings indicate that most teachers use AI assistants occasionally or are only beginning to consider their use, while regular and systematic implementation remains limited. A positive relationship was observed between the level of digital literacy and the frequency of AI use, whereas the most commonly identified barriers include insufficient knowledge, fear of misuse, and the absence of clear guidelines. Overall, attitudes toward AI are generally positive, particularly among teachers with more experience in using such tools, who highlight time-saving effects and improvements in instructional quality. These findings are consistent with patterns described in current research literature and point to the need for targeted professional training and clearly defined ethical frameworks for the use of AI in education.

AI-Driven Transformation of the Fitness Industry: A Case Study of G&S Premium Gym

2026-01-06T10:46:20+01:00

This paper explores the transformative impact of digital technologies on the fitness industry, focusing specifically on the role of artificial intelligence (AI) in enhancing modern exercise equipment. By analyzing the needs of recreational users and professional athletes, it examines how AI-driven fitness devices optimize personalization, improve performance tracking, and elevate the overall user experience. Using G&S Premium Gym in Bijeljina—the first fitness center in Bosnia and Herzegovina to integrate AI-powered equipment—as a case study, the paper delves into the technical specifications, functionality, and user interaction with these advanced machines. Key findings reveal that AI technologies significantly enhance training efficiency and customization, contributing to measurable improvements in user satisfaction and physical performance. However, challenges persist, particularly regarding technology accessibility, user digital literacy, and data privacy concerns. The research highlights the potential for AI to redefine standards in recreation while addressing these challenges. Recommendations for future research and implementation emphasize the importance of affordable, user-friendly AI solutions and improved data security measures.

A Small Language AI Model in the Bosnian Language

2026-01-06T10:55:46+01:00

This study presents the development and evaluation of Mali Mujo, a small-scale language model optimized for the Bosnian language, designed to operate efficiently on devices with limited computational resources. Leveraging the TinyLlama architecture, the model demonstrates the feasibility of deploying natural language processing (NLP) applications in environments with constrained memory and processing capabilities, specifically devices with 1 GB storage and 8 GB RAM. The system integrates Langchain agents and the DuckDuckGo API to enable real-time information retrieval, enhancing the model’s responsiveness and accuracy in practical applications. The methodology involved training the TinyLlama model on a curated Bosnian dataset, followed by testing across diverse real-world scenarios in industry and administration. Performance metrics focused on accuracy, response time, and computational efficiency, while additional evaluation considered user experience and adaptability to domain-specific tasks. The results indicate that Mali Mujo delivers rapid and reliable responses to user queries, with significant advantages in speed and resource efficiency compared to larger language models. The model effectively processes administrative requests, generates technical and market-related insights, and supports educational and governmental applications, highlighting its versatility. While small-scale models exhibit lower absolute accuracy than their larger counterparts, the study demonstrates that careful optimization and integration with external APIs can mitigate limitations, providing a balance between performance and accessibility. Furthermore, the model’s design ensures user privacy and low energy consumption, contributing to sustainable and secure AI deployment. Mali Mujo exemplifies the potential of small language models to enhance efficiency, accessibility, and usability in local-language contexts. Its deployment provides a scalable, cost-effective solution for organizations with limited infrastructure, offering opportunities for further enhancement through expanded datasets, multilingual support, adaptive learning, and integration with emerging AI technologies. The findings underscore the practicality of small AI models in bridging the gap between advanced NLP capabilities and resource-constrained environments.

An Evolutionary Overview of Large Language Models: From Statistical Methods to the Transformer Era

2026-01-07T09:25:36+01:00

While the early evolution of large language models (LLMs), including shift from statistical approaches to the Transformer architecture, illustrates their historical impact on the processing of natural language; however, the latest research in neural networks has enabled the faster and more powerful rise of language models grounded in solid theoretical foundations. These advantages, driven by advances in computing systems (e.g., ultra-powerful processing and memory capabilities), enable the development of numerous new models based on new emerging technologies such as artificial intelligence (AI). Thus, we provide an evolutionary overview of LLMs involved in the shift from the statistical to deep learning approach, highlighting their key stages of development, with a particular focused on concepts such as self-attention, the Transformer architecture, BERT, GPT, DeepSeek, and Claude. Finally, our conclusions present a reference point for future research associated with the emergence of new AI-supported models that are irreversibly transforming the way an increasing number of human activities are performed.

Financial Sustainability of Learning Platforms – Case Study of an E- Learning Project

2026-01-07T09:38:30+01:00

This paper analyzes the potential for investment in the implementation of a multifunctional, sophisticated learning platform through a qualitative research approach based on primary and secondary data collected via surveys and semi-structured interviews with high school students, university students, and employees in need of personalized professional development. The analysis and synthesis of the collected data indicate that the commercialization of the innovative platform is feasible.

Model to improve distance learning system LOOMEN

2026-01-07T09:44:50+01:00

This research study focuses on analyzing the functionalities and potential improvements of the Loomen platform, the most widely used LMS system in Croatian education. The aim of the study was to identify the strengths and weaknesses of the platform and propose solutions that could enhance user experience and improve pedagogical outcomes. The research was conducted during a teaching internship at a high school, where hands-on experience with daily system use was gathered, and through a survey administered to students and teachers. The survey collected both quantitative and qualitative data on usage frequency, technical difficulties, satisfaction with functionalities, and suggestions for improvement. The results indicate that students and teachers value the ability to access teaching materials and submit assignments, but they highlight issues with the user interface, mobile version, and communication tools. Students perceive communication via forums and messaging as outdated and insufficiently engaging, which reduces interaction in the learning process. Based on the collected data, a list of functionalities and identified shortcomings was compiled, followed by proposals for improvement: interface redesign, optimization for mobile devices, introduction of a self-assessment module, and richer multimedia content. The paper concludes by emphasizing that the proposed measures have the potential to increase student motivation and satisfaction, as well as to improve the quality of distance learning.