|
|
 |
Hybrid Event
|
| AI and Large Language Models in Software and Systems Engineering |
P. Rajković, N. Stojiljković (University of Niš, Faculty of Electronic Engineering, Niš, Serbia)
Application of Retrieval‑Augmented Generation Technology in Chatbot Development 
Large language models (LLMs) made a significant step forward and demonstrated substantial progress across general chat platforms and intelligent assistants. However, their deployment in domain‑specific applications is somewhat behind in terms of response quality due to static knowledge boundaries, substantial programming requirements for adaptation, and the potential to generate hallucinated responses. This paper addresses the problem by implementing the Retrieval‑Augmented Generation (RAG) approach and adapting it to serve as a suitable general-purpose help module. Developed system further emphasizes the generation process by supplying the model with appropriate background information retrieved from a semantically indexed knowledge base. As a result, the model’s responses are grounded in concrete, domain‑specific sources instead of depending solely on internal model parameters. The further consequence is fast execution and a low level of hallucination.
|
T. Nikolova, S. Toleva-Stoimenova, S. Syarova (University of Library Studies and Information Technologies, Sofia, Bulgaria)
Bug Prioritization With Large Language Models Using Ontology-Based Prompts
Defect prioritization is crucial for the timely addressing of critical issues from both technical and user perspectives. It is usually performed manually during triage, but it is often prone to errors. In recent years, several studies have focused on the use of AI methods for defect prioritization. However, the proposed approaches mainly treat it as a technical problem. More recent research has focused on LLMs because of the attention to context, but it is limited to certain attributes of the reports. The present paper focuses on exploring the potential of using LLMs for defect prioritization by considering both technical severity and business impact of the issue. To ensure assessment accuracy and minimize hallucinations, they employed ontology-based prompts. A template was created in which the ontology is defined as the context of the prompt, including specified concepts, relations, axioms, and rules. A set of bugs with missing information was provided to investigate whether the LLM can generate new knowledge and draw correct conclusions. The results of this study serve as a basis for developing a Defect Prioritization Application Ontology, which could help optimize triage processes and improve the efficiency of technical teams, as well as enhance user satisfaction.
|
E. Smolić (It from bit d.o.o., Zagreb, Croatia), M. Brčić, L. Hobor, M. Kovač (Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia)
AI-Assisted Unit Test Writing and Test-Driven Code Refactoring: A Case Study
Many software systems originate as prototypes or minimum viable products (MVPs), developed with an emphasis on delivery speed and responsiveness to changing requirements rather than long-term code maintainability. While effective for rapid delivery, this approach can result in codebases that are difficult to modify, also presenting big opportunity cost in the times of AI-assisted or even AI-led programming. In this paper, we present a case study on the utilization of coding models for automated unit test writing and subsequent safe refactoring, with proposed code changes validated by passing tests. The study examines best practices for iteratively generating tests to capture existing system behavior, followed by model-assisted refactoring under developer supervision. We describe how this workflow constrained refactoring changes, the errors and limitations observed in both phases, the efficiency gains achieved, and when manual intervention was necessary. Using this approach, we generated nearly 16,000 lines of reliable unit tests in hours rather than weeks, achieved up to 78% branch coverage in critical modules, and significantly reduced regression risk during large-scale refactoring. These results illustrate software engineering’s shift toward an empirical science, emphasizing data collection and constraining mechanisms that support fast, safe iteration.
|
N. Ranković (Tilburg University, Department of Intelligent Systems, Tilburg, Netherlands), D. Ranković (Union University, Department of Informatics, Belgrade, Serbia)
Agile Issue Triage with Transformers
Issue triage is a central activity in agile software development, where assigning appropriate priority levels directly affects scheduling, resource allocation, and response time. Existing studies on automated issue analysis primarily focus on defect-related tasks, issue type classification, or effort estimation, often relying on lexical representations or shallow embeddings, leaving the role of contextual language models in priority-oriented triage insufficiently examined. This paper evaluates the effectiveness of transformer-based language models for automated priority classification of agile issues using textual descriptions alone. Experiments are conducted on the TAWOS (Tawosi Agile Web-based Open-Source) dataset, comprising real-world issues from multiple open-source Jira repositories, with original priority labels mapped to three ordinal classes. Classical machine learning models based on Term Frequency-Inverse Document Frequency (TF-IDF) representations are used as strong lexical baselines and compared against contextual transformer models, including RoBERTa and DeBERTa-v3, under a unified experimental protocol. Results show that while TF-IDF-based models provide competitive baseline performance, transformer-based models achieve consistently higher Macro-F1 scores and substantially improve detection of High-priority issues. Error analyses further indicate that transformer models reduce extreme misclassifications and better respect the ordinal structure of priority labels by concentrating residual errors between adjacent classes. Generally, the results indicate that modern transformer architectures provide a robust and reliable foundation for decision-support in automated agile issue triage.
|
N. Roso, M. Sužnjević, N. Markuš (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia)
From Games to Robotics: Unifying Human-Centered LLM Integration
Existing Large Language Models (LLMs) research has largely emphasized improving model capabilities and benchmarking performance, while comparatively little attention has been devoted to user-centered factors. These are particularly critical in highly interactive domains, including gaming and robotics, where LLM-driven agents operate in dynamic environments and engage in continuous human–AI interaction. This paper argues that gaming and robotics could form a coupled research ecosystem in which insights, methodologies, and evaluation practices can be shared. We review recent advances in LLM applications across both domains, highlight shared technical and user-centered challenges, and outline a cross-domain perspective in which games serve as scalable testbeds, while robotics ground these insights in physical embodiment and real-world constraints. Finally, we advocate for a unified, user-centered evaluation approach to guide the design of reliable and engaging LLM-driven agents. By bridging gaming and robotics, this work aims to advance systematic, human-centered approaches to LLM integration and accelerate progress toward robust interactive AI systems.
|
M. Grudenić (Smart Aqua d.o.o / UniZG, Zagreb, Croatia)
Toward Assisted Engineering of Urban Water Infrastructure: A Systematic Review of Large Language Models in Urban Drainage and Flood Management
Urban drainage systems are under increasing pressure due to climate-driven pluvial extremes. Conventional engineering workflows based on Geographic Information Systems (GIS) and the Storm Water Management Model (SWMM) are often constrained by labor-intensive data integration, manual calibration, and limited real-time adaptability. The rapid development of Large Language Models (LLMs) introduces a new paradigm, Assisted Engineering, in which intelligent systems actively support engineering analysis and decision-making.
This study presents a systematic literature review of LLM applications in urban drainage and flood management. Following the PRISMA protocol, a structured search of the Scopus database was conducted, complemented by IEEE Xplore and arXiv sources. The reviewed studies are classified into three functional domains: (1) Automated Model Synthesis, (2) Real-Time Decision Support, and (3) Multimodal Spatial Reasoning. The findings indicate that LLM-assisted workflows can reduce model setup time by up to 80% while significantly improving access to complex hydraulic information for non-specialist stakeholders.
|
| Pauza
|
| Radionica |
| "Human-Centered Adoption of LLMs in Industry and Society: Challenges, Opportunities, and Systematic Implications" |
| Pauza |
| Software Quality Programming Paradigms, and Engineering Education |
Z. Stapić (University of Zagreb Faculty of Organization and Informatics, Varaždin, Croatia)
From Code Smells to Curriculum Interventions: Clean Code Deficiencies in Undergraduate Software Engineering Projects
Clean code practices are central to producing maintainable software, yet undergraduate students often deprioritize them under tight deadlines and limited refactoring experience. This paper investigates recurring clean-code deficiencies in undergraduate software engineering projects by analyzing static-analysis findings and their estimated remediation effort. We conducted an empirical study of 44 student project repositories developed between 2022 and 2024, assessed post-course using SonarQube. From the full set of detected issues, we constructed a clean-code–focused subset by filtering for maintainability-oriented code smells (e.g., code hygiene, unused code, control-flow simplification, and related readability/structure problems) and excluding findings outside the scope of clean code (e.g., security vulnerabilities). The resulting dataset includes issue metadata such as type, severity, and technical debt estimates (minutes). Using category-level aggregation, we quantify which clean-code issue groups are the most prevalent and contribute the most remediation effort, and we compare cohort-level patterns across years. Based on these results, we translate the most frequent and the costliest clean-code deficiencies into actionable curricular interventions, including targeted learning activities, assessment criteria, and feedback mechanisms designed to reduce recurring mistakes in future cohorts. The expected contribution is an evidence-based bridge from code-smell analytics to curriculum improvement, providing a replicable approach for using technical-debt signals to prioritize clean-code teaching and assessment.
|
Á. Szauter, N. Pataki (Department of Programming Languages and Compilers, Eötvös Loránd University, Budapest, Hungary)
Rejuvenation of C++ Classes
The C++ language fully supports object-oriented programming, including classes, inheritance, virtual methods, overriding, and the generation of special member functions such as constructors and destructors. Since the C++11 standard, it is possible to explicitly declare special member functions as default using the default keyword, and to mark overridden virtual functions with the override specification. Although the compiler implicitly generates special member functions that are not explicitly implemented, this behavior can be confusing. By using the default keyword, these functions can be declared explicitly. Similarly, users can override virtual functions without using the override keyword, but this can hide subtle bugs, such as accidentally mistyping the name of a function that should be overridden. Using the override specification allows the compiler to catch such mistakes. In this paper, we propose a static analysis approach that can rejuvenate classic C++ classes. We propose a tool that automatically adds missing default member functions to classes and inserts the override specification for functions that override a method in a base class.
|
L. Matovina, M. Kaluža ( Polytechnic of Rijeka, Rijeka, Croatia)
CodeFirst or DataBaseFirst
Managing relational data effectively is essential in modern software development. Object-Relational Mapping (ORM) tools support two main approaches: Code First (CF), where models in code define the database, and Database First (DBF), where an existing schema guides code generation. This study first investigates the general differences between these approaches and evaluates them according to clearly defined criteria, including schema creation, migrations, data seeding, version control, and integration with the application layer. The criteria are contextualized across multiple frameworks, such as EF Core, Hibernate, Django ORM, and Doctrine, to provide a broader understanding of how each approach performs in different technological environments. To further validate the findings and gather practical insights, a survey of experienced practitioners was conducted. Results highlight that CF centralizes control in code, supporting consistent migrations and maintainable development, while DBF offers precise control over physical database structures and performance optimization. These findings provide practical guidance for development teams, assisting in selecting the most suitable approach based on project requirements, maintainability, and scalability.
|
Z. Sirotić (ISTRA TECH d.o.o., Pula, Croatia), S. Sovilj, D. Fonović (Juraj Dobrila University of Pula, Faculty of Engineering, Pula, Croatia), K. Pripužić (University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, Croatia)
Comparison of Multiple Inheritance in Seven OO Programming Languages
Multiple class inheritance has long been unjustly regarded as complex and unnecessary. One of the main reasons is likely that multiple inheritance in C++ is relatively poorly implemented. Specifically, multiple inheritance in C++ was introduced later (1989), rather than from the beginning. Based on such poor experiences, the designers of the OOPL languages Java (and later C#) decided not to support multiple class inheritance. For example, in the Java community, it was long claimed that multiple inheritance is bad, and that inner classes (especially anonymous inner classes) are all that's needed. However, in Java 8, they began moving toward partial support for multiple class inheritance by introducing default methods in interfaces, which previously could not contain code.
The Eiffel language has supported multiple inheritance from the start (1986) and is considered to support it better than any other OOPL. The Scala programming language also supports multiple inheritance almost fully (through so-called traits). This paper analyzes how multiple inheritance is supported (fully or partially) in seven OOPL languages (listed by year of appearance): C++, Eiffel, Java, C#, Scala, Kotlin, Swift.
|
M. Benkus, L. Masnec, Z. Stapić (University of Zagreb, Faculty of Organization and Informatics, Varaždin, Croatia)
AI Inside the Classroom: A Modern Approach to Teaching AI Skills with Azure and Multi-Agent Concepts
Accelerated developments in artificial intelligence (AI) are reshaping the competency requirements of the Information technology (IT) industry, making the integration of AI content into higher education important for developing market-relevant software engineers. This paper proposes a framework applicable to programs aiming to integrate AI competencies. Within the Joint Creative Classroom (JCC) model of collaboration between industry and academia, three AI modules were introduced to modernize the curriculum in accordance with current labor market needs. Accordingly, the research question is: How do AI competencies acquired through the expanded curriculum contribute to the professional relevance and employability of students in the IT industry? Methodologically, the paper relies on the expert review method through semi-structured interviews with three industry experts experienced in software development and candidate selection. The evaluation includes an assessment of the importance of AI topics for employment and an analysis of their contribution to professional skills. The results suggest that AI content effectively bridges the gap between traditional academic learning outcomes and labor market demands. Expert evaluation provides preliminary insight into the critical importance of curriculum modernization for enhancing the employability and professional relevance of future software engineers.
|
A. Lokner Lađević, M. Žagar, D. Delija, G. Sirovatka (Zagreb University of Applied Sciences, Zagreb, Croatia)
Automating Filesystem Artifact Analysis in Linux Environments Using Python
Manual execution of command-line analysis tools can be inefficient and difficult to maintain when applied to complex data processing workflows. This paper presents a Python-based automation approach for filesystem artifact analysis in Linux environments, focusing on improving maintainability, readability, and structured reporting. The proposed solution integrates existing low-level analysis tools into modular Python scripts that automate file enumeration, metadata extraction, and artifact processing.
The implementation demonstrates three automated workflows: identification and extraction of deleted filesystem entries, detection and classification of NTFS Alternate Data Streams, and parsing of structured system data files to extract user-related information. The system is designed using object-oriented principles and provides structured output in JSON, CSV, and human-readable text formats.
Experimental evaluation shows that the Python-based approach reduces manual effort, improves reproducibility, and simplifies result interpretation compared to traditional shell-based scripting. The results indicate that Python offers a practical and scalable foundation for automating complex analysis workflows in Linux-based systems, particularly when structured data processing and reporting are required.
|
| Architectures, Integration, and Lifecycle Management of Complex Systems |
I. Iliev, K. Rasheva-Yordanova, D. Borissova (University of Library Studies and Information Technologies, Sofia, Bulgaria)
An Integrated Model for Optimizing International Postal Addressing and Delivery Services
The rapid growth of cross border electronic commerce has increased the complexity of international postal addressing and delivery services. Many delivery systems are affected by unstructured or ambiguous addresses, geolocation inaccuracies, regulatory differences, and operational inefficiencies in last mile delivery. Existing research mainly focuses on isolated solutions such as address geocoding or route optimization, without addressing the interdependencies between address management and delivery performance. This paper proposes an integrated model for optimizing international postal addressing and delivery services. The model combines address intelligence, hybrid geospatial validation, delivery optimization, and continuous monitoring within a unified framework. It is designed to operate effectively with incomplete or low quality address data and across diverse regulatory and infrastructure environments. The model is evaluated through scenario based analysis and retrospective validation using documented delivery failures. The validation supports the conceptual applicability and practical feasibility of the proposed approach. The main contribution is a holistic framework that connects address processing with delivery optimization and supports future implementation in global postal networks.
|
J. Matkovic (Faculty of Mechanical Engineering, Computing and Electrical Engineering, University of Mostar, Mostar, Bosnia and Herzegovina)
BPMN-Based Visualization of Microservices Choreography Implemented Using Message-Broker Infrastructure
Contemporary approaches to business process automation are increasingly based on the integration of stand-alone, typically JSON-based microservices, which communicate through message exchange over shared, modern message-broker infrastructure. When business logic is realized through interactions among decoupled and autonomous service components operating over a shared communication medium, this paradigm is referred to as service choreography.
A major limitation of the service choreography approach is the insufficient level of process visibility, both during the early stages of business process modeling and in later phases related to process execution monitoring and troubleshooting. This paper proposes the adoption of the well-established Business Process Model and Notation (BPMN) standard to enhance the visualization of business process definitions when they are modeled from scratch using a choreography-based approach, thereby improving applicability in the initial phases of business process modeling.
The paper first examines the inherent limitations of service choreography as well as the fundamental characteristics of the BPMN standard. Subsequently, it presents a proposal for employing BPMN in this context, along with an analysis of its complexity, limitations, and potential benefits.
|
L. Borozan, B. Borozan, D. Ševerdija (School of Applied Mathematics and Informatics, University of Osijek, Osijek, Croatia)
Efficient Checkpointing via Object Serialization in C# Applications
Many modern applications require the ability to suspend execution, create a checkpoint, and later restore their state in order to resume their operation seamlessly. A naive approach to this problem involves capturing a snapshot of the entire memory allocated to the program and saving it to the persistent storage. However, as many applications require maintaining multiple checkpoints, the memory requirements of such an approach render it impractical in realworld scenarios. From the perspective of data reconstruction, program state often contains a significant amount of redundant or runtime-specific information which can be omitted from checkpoint creation and reconstructed when needed. Most modern programing languages are object-oriented, thus program state is encapsulated within objects that can be individually serialized focusing on reducing storage overhead. In this paper we present an efficient C# framework which enables serialization and deserialization of arbitrary objects with the purpose of storing and restoring program state via checkpoints. These checkpoints are formatted to be human-readable and editable using standard text editors. The framework is subject to numerous tests demonstrating its memory efficiency, fast running times, and the ease of integration into existing C# projects.
|
A. Baumgartner, Č. Livada, T. Galba (Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, Osijek, Croatia)
Engineering and Deploying the HoloDent3D Software System Under Real-World Constraints
Modern interactive systems based on advanced visualization technologies promise improved inspection and interaction with complex data, yet their behavior under real-world deployment conditions is often insufficiently understood. While the HoloDent3D system was previously introduced as a conceptual solution for holographic dental visualization, its practical deployment exposes a range of software and systems engineering challenges. This paper presents an application-first engineering analysis of deploying the oloDent3D software system in realistic operational settings. The study focuses on system-level concerns including latency behavior under realistic workloads, resource contention on heterogeneous hardware, runtime configuration trade-offs, and integration with existing software environments. Rather than emphasizing conceptual design, the analysis examines how engineering decisions manifest during deployment and sustained operation, revealing failure modes and performance limitations that are not observable in offline or prototype-centric evaluations. Based on deployment observation, resource monitoring, and operational experience, the paper derives practical engineering lessons and design guidelines relevant to the development, operation, and evolution of complex interactive software systems. The results highlight the importance of deployment-aware engineering practices for achieving performance, reliability, and maintainability in real-world environments.
|
A. Daskalov, A. Kulakov (Faculty of Computer Science and Engineering, University of Ss. Cyril and Methodius, Skopje, Macedon, Skopje, Macedonia), I. Pires (Escola Superior de Tecnologia e Gestão de Águeda, Universidade de Aveiro, Aveiro, Portugal and Insti, Aveiro, Portugal), S. Panev (Faculty of Computer Science, International Slavic University “G. R. Derzhavin”, Sv. Nikole, R. Maced, Sv. Nikole, Macedonia), E. Zdravevski (Faculty of Computer Science and Engineering, University of Ss. Cyril and Methodius, Skopje, Macedon, Skopje, Macedonia)
Evaluation of algorithms for data-driven recommendation systems
This article evaluates the effectiveness of various recommendation system architectures, including popularity-based, collaborative filtering, content-based, and hybrid approaches. The content-based systems include an LLM variant utilizing multiple models and prompts, as well as a traditional NLP variant employing methods such as one-hot encoding, Word2Vec, and TF-IDF. Experimental results highlight the strengths and limitations of each approach, revealing that collaborative filtering outperforms others in most scenarios. In cases with sparse data, popularity-based approaches prove most effective. These findings provide actionable insights into selecting appropriate architectures for recommendation systems under different data constraints.
|
| Pauza |
| Security, Reliability, and Embedded & Distributed Systems |
S. Petrović, M. Tomić (University of Rijeka, Faculty of Engineering, Rijeka, Croatia)
Network Protection with Next Generation Firewalls
The rapid growth of encrypted traffic, cloud services, and application-layer attacks has significantly reduced the effectiveness of traditional firewalls, creating challenges in protecting enterprise networks. This paper investigates whether next generation firewalls (NGFWs) can provide effective, unified protection against contemporary threats. A practical, experimentally validated evaluation of an NGFW deployment is presented, integrating deep packet inspection, intrusion prevention, SSL inspection, user-based access control, and SD-WAN within a single security platform. The methodology is based on a realistic laboratory environment using FortiGate devices to emulate a segmented enterprise network with internal services, remote users, and Internet connectivity. Security policies and services were systematically configured, followed by controlled tests including malware delivery, intrusion attempts, encrypted traffic inspection, and failover scenarios. The results show that the NGFW successfully detected and blocked malicious activity across the tested scenarios, identified multiple simulated intrusion attempts, enforced user-level access control, and provided visibility into network traffic, while full SSL inspection highlighted a trade-off between security, performance, and user experience.
|
D. Sindicic, N. Momcilovic (APIS IT d.o.o., Zagreb, Croatia)
Detecting Remote Code Execution Attacks in Kubernetes Clusters Using an Autoencoder
Applications running on Kubernetes platforms are common targets for cyberattacks. Because Kubernetes can host almost any application environment, its widespread adoption makes it an attractive target for attackers. Kubernetes, the most popular container orchestration system, consists of worker nodes that run applications in isolated environments called containers. Despite this isolation, containers may still be vulnerable to known attack vectors. One such vector is remote code execution (RCE), which allows an attacker to execute malicious commands on a compromised system.
In this paper, we present a command anomaly detection approach based on an autoencoder neural network. The autoencoder is trained on normal commands and detects anomalies as outliers from typical command patterns. We use eBPF and the exec system call to construct a dataset and to capture command executions in real time during evaluation. Experimental results demonstrate that the proposed method successfully detects known attack vectors, including remote code execution attacks.
|
L. Lamani, E. Leka, G. Rexha (Polytechnic University of Tirana, Tirane, Albania), A. Aliti (Mother Teresa University, Skopje, Macedonia)
A Secure and Resilient Framework for Wildfire Management via UAVs and Blockchain Federated Learning
The increasing frequency and severity of wildfires necessitate intelligent and robust airborne surveillance systems capable of functioning in dynamic, lowinfrastructure areas. Unmanned aerial vehicles (UAVs) have become critical for wildfire detection, monitoring, and damage assessment; yet, centralized coordination methods suffer from scalability challenges, sluggish decision-making, and vulnerability to cyber threats. This study presents a federated learning architecture combined with blockchain technology to improve UAV-assisted wildfire management. Federated learning allows UAVs to collectively train models for wildfire detection and severity assessment without sharing raw sensor data, protecting privacy and data location. Blockchain creates a decentralized foundation for trust by utilizing smart contracts to monitor UAV IDs, record model revisions, and enforce collaborative standards. Time-stamped ledger entries and consensus-driven validation help mitigate identity-based attacks such as spoofing and replay attempts, but they do not inherently prevent adversarial model poisoning, which requires additional defenses such as robust aggregation and anomaly detection. The system is designed for high-mobility aerial networks operating in beyond-5G/6G environments and supporting near-real-time wildfire monitoring. The suggested architecture combines federated learning with blockchain to enhance resilience, dependability, and scalability, providing a secure and scalable reference architecture for decentralized, mission-critical wildfire management.
|
F. Krapić, M. Tomić (University of Rijeka, Faculty of Engineering, Rijeka, Croatia)
Fitness Tracker Connected to the Internet Using a Wi-Fi Network
Many open-source fitness trackers rely on Bluetooth Low Energy and a smartphone gateway, which limits standalone operation and complicates continuous data upload. This paper presents a wearable activity tracker prototype that connects directly to Wi-Fi 6 and securely publishes activity metrics to a remote service using MQTT over TLS. The system is built on the Nordic nRF7002DK platform, combining an nRF5340 dual-core microcontroller with an nRF7002 Wi-Fi 6 radio, and it runs Zephyr RTOS via the nRF Connect SDK. Motion is captured with an MMA7361L three-axis analog accelerometer and processed on-device using a step detector with high-pass filtering, dynamic threshold adaptation, and simple activity classification from cadence and estimated speed. A round TFT display with capacitive touch provides a local user interface implemented using the Light and Versatile Graphics Library (LVGL). End-to-end functionality is validated by monitoring encrypted MQTT messages with standard Mosquitto client tools and by exercising a bidirectional control topic for basic commands. The prototype demonstrates a modular architecture for phone-independent wearable telemetry and provides a baseline for future work on energy optimization and improved motion analysis.
|
K. Paić, M. Tomić (University of Rijeka, Faculty of Engineering, Rijeka, Croatia)
GNSS Tracking on the nRF7002 DK
Global Navigation Satellite System (GNSS) receivers are widely deployed in consumer and industrial electronics, motivating compact and network-connected tracking solutions. This work presents the design and implementation of an embedded GNSS tracking device based on the nRF7002 DK platform and the Zephyr real-time operating system. The proposed system acquires raw GNSS output, interprets standard NMEA sentences, and extracts positioning information through an on-device processing pipeline. To enable remote monitoring, the derived location data are transmitted over Wi-Fi - provided by the nRF7002 DK - to a real-time location service for visualization. The resulting prototype demonstrates an end-to-end architecture from GNSS data acquisition to cloud-connected position reporting, and it confirms the suitability of the nRF platform and Zephyr RTOS for developing connected embedded positioning applications.
|
I. Ljubi, M. Vuković (University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, Croatia, Zagreb, Croatia)
Representation Analysis of croBERT for Croatian Disinformation Detection
Transformer-based language models have substantially advanced NLP for low-resource and morphologically rich languages, yet performance gains alone offer limited insight into the linguistic mechanisms underlying their predictions. This study examines how the monolingual Croatian transformer model croBERT encodes disinformation-relevant signals in social-media discourse. Moving beyond aggregate metrics, we perform a representation-level analysis to identify linguistic and stylistic features that drive class discrimination. Using a curated corpus of Croatian user-generated comments annotated for disinformation, we compare feature-based and fine-tuned croBERT models through layer-wise probing, embedding-space analysis, and token-level attribution. Our results show that fine-tuning reshapes representational geometry to improve class separability while reducing reliance on surface lexical cues, with discriminative signals emerging primarily in intermediate layers. These findings deepen understanding of transformer behavior in morphologically rich languages and offer practical guidance for deploying contextual models in disinformation detection.
|
|
Basic information:
Chairs:
Tihana Galinac Grbac (Croatia), Darko Huljenić (Croatia)
Program Committee:
Stipo Čelar (Croatia), Andrej Grgurić (Croatia), Igor Ljubi (Croatia), Mladen Sokele (Croatia), Željka Tomasović (Croatia), Mladen Tomić (Croatia), Linda Vicković (Croatia)
Registration / Fees:
|
REGISTRATION / FEES
|
Price in EUR
|
EARLY BIRD
Up to 15 May 2026 |
REGULAR
From 16 May 2026 |
| IEEE members |
297 |
324 |
| MIPRO members |
297 |
324 |
| Students (undergraduate and graduate), primary and secondary school teachers |
165 |
180 |
| Others |
330 |
360 |
The student discount doesn't apply to PhD students.
NOTE FOR AUTHORS: In order to have your paper published, it is required that you pay at least one registration fee for each paper. Authors of 2 or more papers are entitled to a 10% discount.
Contact:
Tihana Galinac Grbac
Juraj Dobrila University of Pula
Faculty of Engineering
Zagrebacka 30
HR-52100 Pula, Croatia
GSM: +385 99 3820 750
E-mail: tgalinac@unipu.hr
The best papers will get a special award.
Accepted papers will be published in the ISSN registered conference proceedings. Papers in English presented at the conference will be submitted for inclusion in the IEEE Xplore Digital Library.
Location:
 Opatija is the leading seaside resort of the Eastern Adriatic and one of the most famous tourist destinations on the Mediterranean. With its aristocratic architecture and style, Opatija has been attracting artists, kings, politicians, scientists, sportsmen, as well as business people, bankers and managers for more than 180 years.
The tourist offer in Opatija includes a vast number of hotels, excellent restaurants, entertainment venues, art festivals, superb modern and classical music concerts, beaches and swimming pools – this city satisfies all wishes and demands.
Opatija, the Queen of the Adriatic, is also one of the most prominent congress cities in the Mediterranean, particularly important for its ICT conventions, one of which is MIPRO, which has been held in Opatija since 1979, and attracts more than a thousand participants from over forty countries. These conventions promote Opatija as one of the most desirable technological, business, educational and scientific centers in South-eastern Europe and the European Union in general.
For more details, please visit www.opatija.hr and visitopatija.com.
|
|