Abstract: The subject of the research is the distributed training of neural networks in heterogeneous computing clusters. In such systems, nodes have different performance levels. Using synchronous communication leads to the downtime of powerful machines: they sit idle waiting for results from weaker devices. The asynchronous approach provides an alternative where each node updates the global model independently. The objective of this work is the development and evaluation of a software architecture for asynchronous training. The key difficulty is the stale gradients effect. Fast nodes update the global parameters multiple times while a slow node computes a single iteration. As a result, the server receives outdated gradients. The paper describes the design of network communication to minimize delays and compensate for computing speed differences without strict synchronization. The methodological basis is the asynchronous stochastic gradient descent (ASGD) algorithm. Time and accuracy evaluation experiments were performed on the CIFAR-10 dataset. The architecture is implemented in Go (server) and Python (clients). The scientific novelty consists in creating a hybrid architecture with a central Go server and Python clients. The server processes requests in parallel without queues. Applying the Protocol Buffers binary format and the gRPC protocol reduces the transferred data volume and accelerates network exchange compared to text formats. Research conclusions: transitioning to an asynchronous protocol eliminates computing node downtime. Stable convergence requires modifying basic training parameters. The Adam optimizer is unstable under network delays due to momentum accumulation. Classical stochastic gradient descent provides more reliable results. There is a slight accuracy drop due to stale weights, but adding new nodes mitigates this drawback. Experiments confirmed the hypothesis of a regularizing effect: aggregating multiple independent gradients smooths the loss function curve. Future work involves finding mathematical methods to automatically reduce the weight of heavily outdated gradients.

Keywords: deep learning, distributed computing, asynchrony, heterogeneous systems, gradient aggregation, client-server architecture, neural networks, parallel processing, golang, gRPC

Abstract: The subject of the research is the task of automated classification of text inquiries from citizens, which contain various problems and requests directed to state and municipal authorities. The focus is on identifying opportunities for effective and practical application of machine learning algorithms for the automatic processing and categorization of these inquiries. The study covers the complete analysis cycle: from the preprocessing of texts and their representation in numerical form to the training of classification models and assessing the quality of their performance. The specifics of the subject area are also studied in detail, including the characteristics of the vocabulary used in citizen inquiries, the presence of unstructured data, significant variability in phrasing, and the problem of class imbalance. The work aims to enhance the efficiency of document management systems and the automated processing of texts in government information systems. To achieve the task, machine learning methods were used, with text preprocessing through TF-IDF and training a classifier based on logistic regression. The main conclusions of the conducted research include confirmation of the effectiveness of adapting classical machine learning methods for the automated classification of citizen inquiries, as well as the identification of a significant influence of model parameters on the quality of the final classification. Specifically, it was shown that using the parameter "class_weight" = "balanced" allows accounting for class imbalance without compromising the model's stability. It was also established that the choice of optimization algorithm—"lbfgs," "newton-cg," "sag," or "saga"—affects the accuracy and speed of learning. A practical approach to adjusting class weights is proposed, ensuring improved results when working with unstructured texts. The analysis conducted showed that even with relatively simple models, such as multinomial logistic regression, it is possible to achieve high accuracy with proper data preparation and correct parameter tuning. Thus, the application of automatic methods for processing textual information can significantly improve the quality and efficiency of the categorization of inquiries, accelerate the analysis process, and simplify the implementation of such solutions in electronic document management practices.

Keywords: Text classification, Machine learning, Socio-political communications, Logistic regression, Automated text analysis, Natural language processing, Text analytics, Government information systems, Text vectorization, Class imbalance

Abstract: In modern natural language processing systems (semantic search, clustering, recommendation systems, RAG systems), the key operation is obtaining dense vector representations of texts—embeddings. The quality of Russian embeddings is primarily determined by the encoder model used; however, with a large number of requests, the embedding generation stage becomes an architectural "bottleneck" that limits the throughput of the entire system. This article explores the possibility of removing this limitation by designing and implementing an embedding generation service in the system programming language Rust, using the dynamic batching algorithm and the ONNX Runtime engine for running machine learning models, as well as comparing its performance metrics with a classical solution in Python. For experimental comparison, two versions of the service were designed and implemented: in Python (using the FastAPI, PyTorch, and Transformers frameworks) and in Rust (using the Axum and ort frameworks). Load testing was conducted to simulate a short-term stress load on the RAG system, analyzing the main performance metrics of the systems using the Locust tool. The results of the load testing show that the Rust service demonstrates multiple advantages over the Python version in all scenarios across all assessed metrics (throughput, latency at the 50th and 99th percentiles, and average RAM consumption). The article analyzes the reasons for the superiority of the developed service in Rust: conversion of the model to the ONNX format, absence of global interpreter lock, true multithreading, application of the dynamic batching algorithm, and efficient memory management without a garbage collector. The results of the study prove the suitability of the system programming language Rust for building high-performance embedding generation services and confirm its significant advantages over the classical Python solution under limited computational resources.

Keywords: text embedding, model inference, dynamic batching, high-load services, natural language processing, perfomance, throughput, latency, Rust, programming language

Abstract: The subject of the study is the capabilities of modern open large language models (LLM) in solving tasks in professional subject areas and their suitability for practical application in industries. The paper presents the results of comparative testing of 15 open-source models with the number of parameters ranging from 8 to 120 billion. The testing was conducted on an original Russian-language benchmark covering 16 industry domains: energy, law, finance, healthcare clinical processes, therapeutic practice, pharmaceuticals, construction, DevOps, logistics, marketing, HR, labor protection, agriculture, sales, procurement/tenders, and manufacturing engineering. For each domain, a set of 90 questions was formed, distributed across three levels of difficulty; the total volume of the benchmark is 1,440 questions. The study examines the models' resilience to increasing task complexity, their domain specialization, and sensitivity to generation parameters. The research method involves comparative testing of models on a single set of questions. Evaluation of the answers was performed in two stages: automatic assessment by three independent judging models using the LLM-as-a-Judge methodology (based on the criterion of semantic correspondence to reference answers) and independent expert evaluation by practitioner specialists with subsequent verification of consistency. The scientific novelty of the work lies in the creation of an original Russian-language benchmark comprising 1,440 questions of three difficulty levels for 16 professional subject areas (from energy and healthcare to DevOps and manufacturing engineering), as well as in the two-stage verification of results, combining automatic assessment with independent expert validation. An analysis of the consistency of the judges' models' evaluations and their comparison with assessments from practitioner specialists was conducted. The testing results indicate that modern models with 27–31 billion parameters, including Mixture-of-Experts architectures and models with FP8 optimization, provide quality of answers comparable to models with 120 billion parameters or exceed them while consuming significantly fewer computational resources. A recommendation system for selecting open language models for industry applications is proposed.

Keywords: large language models, LLM, benchmark, LLM-as-a-Judge, open-source, domain knowledge, professional evaluation, multi-judge evaluation, expert validation, comparative testing

Abstract: The paper presents a conceptual architecture for continuous business process reengineering based on the integration of Agentic AI and Digital Twin of Organization (DTO) technologies. The relevance of the study is driven by the fundamental inadequacy of the classical BPR paradigm under conditions of high environmental volatility in 2025–2026, where market windows shrink to weeks while traditional reengineering projects span one to five years. The object of the study is the operational management processes of digital enterprises. The subject of the study is the methods and architectural approaches to automated continuous transformation of business processes. A model of Autonomous Business Process Reengineering (ABPR) is proposed, comprising a five-layer technological architecture, a multi-agent system with four differentiated functional roles – monitoring, diagnostics, solution synthesis, and deployment – a taxonomy of process changes by levels of AI autonomy, and an embedded Human-in-the-loop control mechanism. The Digital Twin of Organization serves as a bidirectional infrastructure element that continuously receives operational data from ERP, CRM, and WMS systems and returns verified process changes back into production environments. Agent orchestration is implemented via the ReAct framework extended with long-term vector memory and asynchronous inter-agent communication. Validation through simulation modeling across two industry scenarios – last-mile logistics and fintech retail lending – demonstrates a reduction of the full reengineering cycle by a factor of 8–15 compared to the traditional project-based approach, alongside a projected improvement in change quality of 17–23 percentage points. The scientific novelty lies in the development of the first systematic integrated ABPR architecture in which Agentic AI acts as a subject of process governance rather than a task automation tool, combined with an original taxonomy of process changes and the concept of a bidirectional Digital Twin of Organization as an active management element.

Keywords: business process reengineering, agent artificial intelligence, digital twin of an organization, process Mining, business process management, process Intelligence, multi-agent systems, human-in-the-loop, low-code platforms, digital transformation of the enterprise

Abstract: The subject is a temporal graph in the Neo4j DBMS designed to store interval facts, Allen temporal dependencies, and derived objects of a temporal knowledge base. The aim of the study is to compare compact and extended storage structures for interval relations when temporal queries are executed and events are received in batches. Two graph schemas are considered. In the first schema, intervals are stored as properties of contacts. In the second schema, contacts and intervals are represented by separate nodes, computed Allen relations are stored as typed relationships. Relational schemas, including a table of temporal dependencies, are used as a control model. Special attention is given to the relationship between the storage method, query execution time, graph size, and construction of derived objects in the temporal knowledge base. The research methods include processing open data of contacts between students, transforming discrete observations into intervals, building graph and relational structures, checking local consistency of Allen relations, executing control queries, and processing event batches. This processing records the formation of derived knowledge-base objects and decision-support situations. The novelty of the study lies in the experimental comparison of compact and extended representations of Allen relations in Neo4j and in the relational model. The effect of precomputed temporal links on different classes of queries is examined separately. The results show that the extended structure accelerates queries over specific relations, but increases graph density and slows down multi-step traversals. The compact structure requires less storage and is suitable for filtering, aggregation, and short chains. The extended structure is useful for repeated queries over Allen relations, temporal rules, and tracing derived objects back to source facts. The storage structure should be selected according to the query class, the acceptable number of stored links, the temporal horizon, and the transaction size.

Keywords: temporal database, temporal knowledge base, Allen’s interval algebra, graph structure, Neo4j graph DBMS, Cypher, temporal rules, decision-making, incremental processing, Python