This research presents a comparative study of rule-based and machine learning-based approaches for detecting anomalous authentication activities. Rule-based detectors are evaluated against an unsupervised anomaly detector trained on normal user behavior, using the LANL dataset expanded with realistic synthetic attacks. Thresholds used by all detectors are calibrated on an evaluation set to meet fixed false-positive budgets. Results are reported using eventlevel and burst-level metrics. The results show that rule-based approaches perform strongly on high-rate attacks, while machine learning approaches are effective for low-rate, stealthy activity.

Microservices systems often face performance issues when workloads fluctuate, and services degrade over time. Traditional load balancing methods such as Round Robin or Latency-Aware routing do not adapt to changing conditions, which can lead to higher latency and increased error rates. This paper evaluates adaptive decision-making algorithms for request routing, including Deep Q-Network (DQN), Upper Confidence Bound (UCB), Thompson Sampling, and traditional heuristics. Experiments were executed on a production-scale cloud environment (Runpod, 16 vCPUs, 128 GB RAM) for 4 hours per algorithm with 50 concurrent users, generating more than 600,000 requests per experiment. Results show that contextual bandit algorithms significantly outperform deep reinforcement learning. UCB achieved a 0.097 % error rate and a median latency of 220 ms, compared to DQN which produced an 11.32 % error rate and instability during training. Latency-Aware routing performed well but could not match the adaptability of contextual bandits. These findings demonstrate that simpler learning algorithms such as UCB and Thompson Sampling provide faster adaptation, lower error rates, and better stability than deep RL approaches in microservices routing tasks.

Banking systems nowadays handle millions of transactions every day, where speed matters most when the system must detect fraud. Traditional batch-processing systems introduce delays because data is being processed at scheduled intervals. Event-driven architecture handles each transaction at the moment it appears; therefore, the system can react almost immediately. This paper compares event-driven and batchprocessing architectures using simulated banking transactions. The results show that event-driven processing significantly reduces latency and enables earlier fraud detection, while batch processing still works well for non-critical jobs, such as periodic user profiling.

Production CRM systems increasingly use large language models, yet typical Retrieval-Augmented Generation (RAG) implementations suffer from knowledge staleness due to 5–10 min batch processing cycles. This paper presents a streaming RAG architecture for business CRM applications that provides real-time knowledge updates with average document-to-query propagation latency of 3.1 s and strong retrieval quality. The event-driven system uses Apache Kafka for document ingestion, Rust microservices for embedding generation, PostgreSQL with pgvector for vector storage, and GPT-4 for response generation. On 62 insurance policy documents from 20 users and 102 test queries, mean document-to-query propagation latency was $3.1 ~\mathrm{s}, 75-150 \times$ faster than batch processing, with retrieval quality metrics of Precision@5 = 0.398, MRR $=0.938$, and NDCG${@} 10=0.942$ consistent with values reported in prior literature. Additional load testing with simulated users verified production-grade performance stability (P95 latency $<10.33 ~\mathrm{s}$), suggesting that streaming designs may mitigate the knowledge-currency vs. system performance trade-off in production CRM applications.

This paper benchmarks LLM-generated synthetic data for fine-tuning RoBERTa-base on two GLUE tasks (SST2 sentiment classification and MRPC paraphrase detection) under a low-resource setting with 1,000 real training examples per task. Real-only, synthetic-only, and hybrid (1 k real + 1 k synthetic) regimes are compared using data from eleven contemporary LLMs. Results show that synthetic-only training remains below real-only baselines, but hybrid training consistently improves performance: on SST-2, the best hybrid configuration nearly matches doubling the real data, while on MRPC gains are smaller but positive. LLM-generated text is most effective as a supplement rather than a replacement for human-labeled data.

This study explores how the choice of database influences performance, modularity, and extensibility in monolithic and microservices software architectures. These software architectures are tested in combination with relational database MySQL and document-based database MongoDB. Apache JMeter is used for automated load testing. Relational databases consistently deliver better performance for structured transactions, while document-based solutions offer greater flexibility and extensibility in distributed systems. In addition, the analysis shows that the choice of database may have more effect on the total performance of a microservices architecture than the inherent overhead of the architecture itself. The findings highlight the critical trade-offs between performance and flexibility, emphasizing the importance of strategic database selection.

Despite the fact that technology is improving day by day and that the medical devices (MDs) are being constantly upgraded, their malfunction is not a rare occurrence. The aim of this research is to develop an expert system that can predict whether the device will satisfy functional and safety requirements during a regular inspection. This expert system can be seen as part of Industry 4.0 that is revolutionizing medical device management. In order to develop the system, five machine learning algorithms that are representative of each classifier group, were used: (1) Random Forest, (2) Decision Tree, (3) Support Vector Machine, (4) Naive Bayes, (5) k-Nearest Neighbour. The Decision Tree outperformed other classifiers achieving the classification accuracy of 100% with and without attribute selection applied on the dataset. This study showed that machine learning algorithms can be used in order to predict MDs performance and potential failures in order to make the process of maintenance of medical devices more convenient and sophisticated and it is one step in modernizing medical device management systems by utilizing artificial intelligence.

A car price prediction has been a high-interest research area, as it requires noticeable effort and knowledge of the field expert. Considerable number of distinct attributes are examined for the reliable and accurate prediction. To build a model for predicting the price of used cars in Bosnia and Herzegovina, we applied three machine learning techniques (Artificial Neural Network, Support Vector Machine and Random Forest). However, the mentioned techniques were applied to work as an ensemble. The data used for the prediction was collected from the web portal autopijaca.ba using web scraper that was written in PHP programming language. Respective performances of different algorithms were then compared to find one that best suits the available data set. The final prediction model was integrated into Java application. Furthermore, the model was evaluated using test data and the accuracy of 87.38% was obtained.