ABSTRACT. This paper presents a systematic comparison of six deep learning architectures – ResNet-50, DenseNet-121, EfficientNetV2-S, MobileNetV3-Large, Swin-T, and ConvNeXt-Tiny – for the task of building damage severity classification from unmanned aerial vehicle (UAV) imagery. The classifier under study constitutes the third stage of the cascaded pipeline YOLOv8 → SegFormer → classifier proposed in prior work, where buildings are first detected, then segmented, and finally assigned a damage severity label. To isolate architectural effects from confounding factors, all six models are trained under identical conditions on the RescueNet dataset adapted via per-class connected component analysis, yielding 9 700 building crops distributed across four damage classes. Each classifier receives a 224×224 RGB×mask product, which zeroes background pixels and reproduces the data flow of the production cascade. Models are trained for 30 epochs with AdamW (learning rate 1×10-4, weight decay 1×10-4), cosine annealing with 3-epoch warmup, weighted cross-entropy and seed 42 on a single NVIDIA T4 GPU. EfficientNetV2-S achieves the highest overall quality (accuracy 80.1 %, macro F1 = 0.813, latency 19.1 ms/image), while ConvNeXt-Tiny offers the best speed/accuracy trade-off (macro F1 = 0.795, latency 6.27 ms/image) and lies on the Pareto frontier together with EfficientNetV2-S. Per-class analysis reveals that boundary classes (no_damage, destroyed) are recognised reliably (F1 ≥ 0.82), whereas the two intermediate classes (minor_damage, major_damage) remain challenging (F1 = 0.63–0.76) due to inherent annotation ambiguity. The recommended cascade configuration is EfficientNetV2-S for offline analysis and ConvNeXt-Tiny for on-board deployment, completing the detection–segmentation–classification system.

ABSTRACT. Offshore wind operations require reliable estimates of weather and sea-state windows for installation, inspection, maintenance, and safe access to offshore structures. Many operating-window assessments apply deterministic thresholds to a single metocean source, which can hide uncertainty caused by source disagreement, missing observations, and irregular in-situ measurements. This study develops a reliability-aware big-data workflow for the North Sea offshore wind area during 2020--2025. The workflow combines Copernicus Marine North West Shelf wave reanalysis, ERA5 atmospheric reanalysis, North West Shelf ocean physics reanalysis, and in-situ observations to analyse wave--wind operating-window limits, with current speed retained as hydrodynamic context. The heterogeneous data are harmonised on a 3-hourly reference axis and transformed into safe-time fraction, accepted window duration, seasonal accessibility, source-agreement metrics, and reliability indicators. Reliability is evaluated through data completeness, observation availability, source agreement, and missingness. Reanalysis-only and observation-informed classifications are then compared to identify false-safe and false-unsafe operational periods. The workflow makes the data basis of each operating-window estimate explicit and separates threshold-based accessibility from the reliability of the supporting metocean evidence.

ABSTRACT. A comparative study is carried out of five regression models (Linear Regression, ElasticNet, Bayesian Ridge, Random Forest, SVR) and three reference imputation methods (mean imputation, kNN, MissForest) for recovering missing values in hydrochemical monitoring data of the Kaniv Reservoir (24 observations, six stations, 2021–2025). Total suspended solids, ammonium, nitrite, nitrate, and inorganic phosphorus are taken as target parameters. Model quality is assessed by leave-one-out cross-validation with 95 % bootstrap confidence intervals and the Wilcoxon signed-rank test on paired residuals. An additional verification is performed on an independent set of real measured values. The selected models (ElasticNet for total suspended solids and nitrate, Random Forest for ammonium, nitrite, and inorganic phosphorus) deliver an MAE improvement of 3– 42 % over mean imputation (statistically significant at the uncorrected level α=0.05 for ammonium and nitrite). Nitrite is analysed separately because its concentration lies in the sub-milligram range, which makes relative metrics (MAPE) for this parameter less reliable and demands cautious interpretation.

ABSTRACT. The article examines the transformation of the educational process under the influence of generative artificial intelligence through the lens of the Three-Subject Didactics 2.0 framework. The authors propose an intelligent testing system implemented on the basis of the CrewAI technology. The system is designed as a multi-agent architecture in which each agent performs narrowly specialized functions, including the verification of student answers, thematic analysis, and the generation of individualized recommendations. The architecture of the multi-agent system is complemented by the NotebookLM platform, which serves as an intelligent source of educational content. The choice of NotebookLM is determined by its architectural feature of source grounding, which ensures that responses are generated exclusively on the basis of materials preselected and uploaded by the instructor, thereby minimizing the risk of hallucinations or inaccurate information. The testing system generates personalized prompts for each incorrect student answer, aimed at addressing identified knowledge gaps. The student, interacting with NotebookLM based on the generated prompts, performs targeted exploration of explanations, analyzes relevant fragments of the provided sources, and develops a refined understanding of the learning material. The use of NotebookLM in this process enables in-depth learning supported by verified sources. The scientific novelty of the study lies in the integration of a multi-agent approach and the Three-Subject Didactics 2.0 conception to create an adaptive, transparent, and verifiable assessment environment. The proposed solution enhances feedback quality, supports personalized learning, and fosters the development of students’ critical thinking skills. It may also be applied in higher education systems to support individualized learning trajectories.

ABSTRACT. The growing use of large language models in IT education introduces a structural quality risk: AI-generated instructional sequences may be statistically plausible yet logically incoherent, guiding learners through steps that are not substantively grounded in their verified competency state. This paper proposes a Binder-based framework grounded in connexive engineering to address this problem in AI-supported IT education. A Binder is defined as a logic-governed mediation unit that maps a learner’s verified competency state to a logically admissible next instructional step and constrains AI guidance through substantive derivation rather than probabilistic suggestion alone. The framework is operationalized through three mechanisms: implication validation, connexive diagnostic analysis, and logical stitching. Binder behavior is further mapped onto a tri-subjective model of AI participation, denoted as S0–S3, in which coherence management develops from teacher-defined rules to AI-supported validation under human oversight. Using a design-science methodology, the paper presents a specification of the framework and demonstrates its application through a worked case in database normalization, a domain characterized by strict prerequisite dependencies that are representative of professional IT training. The case shows how Binder-based sequencing can detect prerequisite bypasses, structure progression across competence levels, and support accountable AI-mediated instructional guidance. The paper contributes a research-in-progress design artifact that operationalizes connexive logic as a set of architectural constraints for AI-supported professional IT education.

ABSTRACT. The algorithm assigns researchers to one or more research groups according to the Australian and New Zealand Standard Research Classification (ANZSRC-2020) is proposed. In terms of the ANZSRC-2020, a research group refers to a specific field of research. For decision making the algorithm only processes a short set of arbitrary, self-reported keywords provid-ed by a researcher. These keywords may be listed on a Google Scholar pro-file, for example. The proposed algorithm operates in three stages. In the first stage, the search queries are formed by the initial keywords and their pair-wise combinations. At the second stage, the membership degrees of queries to research groups are computed. The overall distribution of the number of publications over research groups for the last five years is found using the Dimensions API. At the third stage, all the query distributions are averaged, resulting in a one-dimensional vector. We further categorize the minor re-search groups by considering their kinship. We collected a dataset of 2,489 researchers from Dimensions and Google Scholar for tuning the parameters of the algorithm. After the tuning we compared the distributions based on the researchers' interests and their papers. The number of researchers with excel-lent matches is 28%. This is a relatively high rate for such a type of source information, considering that it is based on just two to five keywords. The proposed categorization algorithm may be useful for quickly producing a short list of candidates for reviewer assignment problems and similar tasks. It is also useful for categorizing researchers who have interdisciplinary papers with many co-authors, where it is unclear which keywords of the publication belong to each co-author.

ABSTRACT. The integration of artificial intelligence into educational processes challenges traditional didactic models and necessitates new approaches to structuring learner interaction with AI systems.

This paper introduces the concept of White Prompting within the framework of Three-Subject Didactics 2.0, where learning is understood as a triadic interaction between student, teacher, and artificial intelligence. White prompts are defined as controlled interaction patterns that constrain the use of AI in order to preserve and stimulate learners’ cognitive activity rather than replace it.

The study proposes a methodological approach to the design of white prompts and conceptualizes their role in regulating the level of cognitive delegation to AI. Two application cases are presented in the domains of mathematics education and language learning, illustrating how structured prompting can support different types of cognitive engagement.

ABSTRACT. This paper presents a lightweight browser-based 3D engine designed to support the development and execution of interactive educational scenarios without requiring native runtime installation. Existing approaches to delivering 3D learning content through the browser force developers to choose between heavyweight exported runtimes with long loading times and low-level graphics libraries that demand specialized rendering expertise. The proposed solution provides a high-level component model over a hidden rendering backend, enabling educational platform developers to build interactive 3D scenarios with reduced engineering effort while maintaining runtime efficiency on consumer hardware. The paper describes the key architectural choices behind the engine and introduces three targeted optimization strategies addressing asset memory consumption, rendering initialization latency, and per-frame computational overhead. The evaluation covered 87 configurations across three benchmark scenes, using three graphics adapters on two machines. GPU-compressed textures reduce the cost of the first rendered frame by 56.7% to 69.5%; shader pre-compilation is effective on the discrete adapters but not on the integrated one; and dirty-flag transform propagation trades a per-phase speed-up for a frame-level cost in the workload measured. The results indicate that the approach is practical for browser-delivered educational 3D content, and that the benefit of a given optimization does not follow a simple low-end to high-end ordering.

ABSTRACT. Background: Combat-related trauma is associated with severe wound contamination, repeated surgical interventions, prolonged hospitalization, and extensive exposure to broad-spectrum antimicrobial therapy, creating favorable conditions for the emergence of multidrug-resistant organisms (MDROs). Modern wartime conditions require adaptive antibiotic stewardship strategies capable of integrating microbiological, clinical, and trauma-related data into structured decision-making systems. Objective: To present the concept and preliminary clinical implementation of the ATAS (AI-Assisted Trauma Antibiotic Stewardship) digital platform designed to support antibiotic stewardship in combat-related trauma. Materials and Methods: This observational study included microbiological and clinical data obtained from 432 patients with combat-related injuries treated between 2022 and 2026. A total of 1,720 bacterial isolates from wound specimens were analyzed. Clinical variables included mechanism of injury, wound characteristics, surgical interventions, intensive care unit stay, previous antibiotic exposure, microbiological findings, and resistance profiles. The ATAS platform was developed as a web-based digital decision-support ecosystem integrating trauma characteristics, microbiology, antimicrobial resistance patterns, and antibiotic exposure trajectories into a structured clinical workflow. Descriptive statistics, temporal trend analysis, and comparative resistance analysis were performed. Continuous variables were expressed as mean ± standard deviation, while categorical variables were presented as frequencies and percentages. Statistical significance was defined as p < 0.05. Results: Gram-negative microorganisms predominated among isolated pathogens, accounting for 71.8% of all isolates. The most frequently identified pathogens were Acinetobacter baumannii (28.4%), Pseudomonas aeruginosa (21.6%), and Klebsiella pneumoniae (18.9%). Carbapenem resistance was identified in 62.7% of A. baumannii isolates and 41.3% of K. pneumoniae isolates. Polymicrobial wound infections were observed in 58.1% of patients. Progressive increase in multidrug-resistant isolates was observed during the study period (p=0.003). Patients requiring ≥3 surgical procedures demonstrated significantly higher prevalence of MDRO-associated infections compared to patients undergoing fewer interventions (74.5% vs 49.2%, p=0.01). The ATAS platform enabled structured integration of microbiological and clinical variables, visualization of resistance-associated trends, and adaptive stewardship-oriented clinical support. Conclusions: The ATAS platform represents a novel AI-assisted digital ecosystem for antibiotic stewardship in combat trauma. Integration of microbiological surveillance, resistance-oriented analytics, and structured clinical support may improve adaptive antibiotic decision-making and facilitate personalized antimicrobial stewardship in complex wartime environments.