Serbian entrepreneurs lack adequate information on tax laws and administrative procedures, which creates various problems for them and their bookkeeping company. This situation is exacerbated by inconsistent tax regulations and complex administrative processes. We developed ValidoAI, an innovative AI-based system with multiple modules addressing these challenges. The system automates tax-related processes and provides personalized guidance through natural language processing (NLP) models. The Tax Regulation Search (TRS) module helps small business owners understand their financial obligations through automated data visualization and analysis. The platform integrates essential functionalities, including employee-level payroll analysis, dynamic cash flow tracking, and structured interpretation of financial reports. The system’s architecture features containerized microservices with specialized preprocessing for Serbian language nuances across both Cyrillic and Latin scripts. System structure and functionality were modeled using the ArchiMate standard, enabling precise representation of business, application, and technology layers. Key functionalities include anomaly detection, supplier similarity identification, and analytical feedback generation through advanced NLP techniques, including word embeddings, vector-based classifications, and clustering algorithms. Testing was conducted through a GPT-powered interface enabling real-time interaction with accounting data. Evaluation demonstrates 92
In recent years, application development has increasingly embraced microservice architectures, with containerization, serving as its foundation. Kubernetes has emerged as the de facto orchestrator of container lifecycle in cloud scale. In this paper, we review novel scheduling solutions proposed by the research community to address issues like resource fragmentation, interference and unexpected evictions, that impact resource utilization and solutions to mitigate environmental impact. We also review strategies to address scheduling complexities in realistic multi-cluster environments. This work will serve as a valuable reference for researchers exploring this emerging field.
This research focuses on C# code smell detection—structural defects that increase technical debt and worsen software maintenance. It aims to compare the effectiveness of fine-tuning with Retrieval-Augmented Generation (RAG) three recent open-source LLM models—DeepSeek-Coder-V2-Lite-Base, LLaMA-3.1-8B-Base, and Qwen2-1.5B-Base—to enhance their versatility for context-aware code smell detection, overcoming limitations of static analysis tools and specialized AI models. The approach relies on parameter-efficient fine-tuning with QLoRA, augmented by RAG, to adapt the models for identifying Complex Method, Complex Conditional, and Feature Envy defects using the DeepLearningSmells dataset sourced from GitHub, maintaining the models’ other capabilities. The research promises to deliver adaptable, AI-driven solutions for improving software maintainability, with empirical validation on a balanced test set demonstrating substantial performance gains over baselines. An ablation study confirms the integral role of RAG and QLoRA in achieving these gains.
Since melanoma is one of the most aggressive types of skin cancer, accurate detection and as soon as possible, by predicting the correct diagnose, is essential to improving patient health. Traditional methods that usually rely on the visual inspection of experts, it can be subjective and time consuming. The goal of this study is to create a deep learning-based system for melanoma classification from dermoscopic images in order to support clinical decision-making, also to improve diagnostic precision. In this study, we propose an approach based on deep learning, especially transfer learning with a ResNet50 architecture to classify melanoma from dermoscopic images. To train and test our idea, we have used the images of ISIC 2020 dataset. To handle the problem of having too few examples in one class and to help the algorithm to perform better on other new data, we use various techniques like data augmentation, oversampling to increase the size of the dataset, in order to balance the classes and make the model more reliable when testing. From the pretrained ResNet50 model, we added configurations like learning rate, weight decay, and how deep the network is. Before training, we have passed all skin images through pre- processing steps, normalization, resizing, hair removal etc., in order to make sure the model gets more clean and better data to train it. In addition, we tried to create a system similar to the first one, using now only dermoscopic images from Albanian patients and train, test a new model based only on Albanian metadata. Results achieved from all experiments show that the proposed methodology achieves high classification accuracy. Metrics that we use for evaluation of the results are accuracy, precision, recall, and F1-score.
Given that digital education highly relies on cybersecurity, ethical hacking has become one of the main disciplines meant to protect all cybernetic devices and platforms. The paper presents a new approach in analyzing IP addresses captured by GlassWire Firewall at the University of Tetova-LMS. These IP addresses become helpful in extracting and interpreting the patterns that denote unusual or potentially malicious activity. This way, the understanding of network security dynamics is gained. In addition, we would be carrying out in-depth log analysis: access and error logs give insight into the vulnerabilities within this landscape. Specific threats will be highlighted in findings and targeted security measures will be proposed to mitigate these risks. Unlike other research works, this study has also executed the ethical hacking techniques, such as penetration testing and vulnerability scanning, which shall enable a closer look into whether LMS’ security framework is flawed or not. These will finally sum up our work with actionable strategies that aim at hardening system defenses. The continuous monitoring and proactive measures have turned out to be critical in taming cyber threats against the studied educational platforms. This research, while it contributes to the scientific field of cybersecurity by taking into consideration the interaction between ethical hacking and educational technology, further aims to develop a set of practical solutions that will be used by educational institutions.
