The presented paper is devoted to the development of a method for identifying and correcting spelling errors in Kazakh texts. In this paper, the study object is methods for more accurate correction of spelling errors in Kazakh texts. The aim of the study is to develop an augmented version of the Damerau-Levenshtein method for correcting spelling errors in Kazakh language texts. Automatic detection and correction of spelling errors have become a default feature in modern text editors for working with text data, in text messaging applications such as chatbots, messengers, etc. However, although this task is well solved in geographically widespread languages, it has not been fully solved in languages with a small audience, such as the Kazakh language. The methods developed so far cannot correct all spelling errors found in Kazakh texts. Therefore, the development of a method with specific algorithms for spelling error correction in Kazakh texts is considered. As a result of the research work, algorithms for correcting errors found in Kazakh language texts were developed, and the developed algorithms were included in the Damerau-Levenshtein method. The experimental testing results of the augmented Damerau- Levenshtein method showed 97.2 % accuracy in correcting specific errors found only in Kazakh words and 92.8 % accuracy in correcting common errors from letter symbols. The standard Damerau-Levenshtein method testing results showed 76.4 % accuracy in correcting specific errors found only in Kazakh words. The results of the tests in correcting common errors from letter symbols with the standard Damerau-Levenshtein were approximately the same with the augmented Damerau-Levenshtein method, the accuracy is 92.2 %. The extent and conditions of practical application of the results are implemented by including them in text editors, messengers, e-mails and similar applications that work with text data.

Federated learning (FL) enables distributed model training in IoT environments while keeping raw data on local devices. However, protecting model-update exchange is difficult on microcontroller-class devices due to strict latency, memory, and energy constraints. Existing studies often evaluate lightweight cryptography outside complete FL pipelines or on more powerful hardware, leaving its practical overhead on MCU-class devices insufficiently explored. This paper presents an end-to-end, hardware-validated secure framework for exchanging model updates in federated learning on resource-constrained IoT microcontrollers. Implemented on ESP32-based edge devices, the framework combines lightweight block ciphers (SPECK, SIMON, and PRESENT), HMAC-SHA256 for integrity verification, and ECDH-HKDF for session-key establishment. The evaluation assessed latency, throughput, RAM/ROM footprint, and energy consumption. Results show that SPECK provides the lowest overhead (0.13 µs/byte, 8.68 MB/s, 138.3 mJ), SIMON offers intermediate performance (0.41 µs/byte, 1.96 MB/s, 184.9 mJ), and PRESENT incurs the highest computational cost (89.37 µs/byte, 0.011 MB/s, 446.2 mJ). In the CICIoT2023 federated intrusion-detection evaluation, the secure model maintained stable convergence and achieved 85.43% accuracy after 20 rounds, remaining close to the centralized baseline. These findings demonstrate the practical feasibility of secure model-update exchange in FL on real IoT microcontrollers and provide hardware-grounded guidance for cipher selection under tight resource budgets.

Federated learning (FL) enables distributed model training in IoT environments while keeping raw data on local devices. However, protecting model-update exchange is difficult on microcontroller-class devices due to strict latency, memory, and energy constraints. Existing studies often evaluate lightweight cryptography outside complete FL pipelines or on more powerful hardware, leaving its practical overhead on MCU-class devices insufficiently explored. This paper presents an end-to-end, hardware-validated secure framework for exchanging model updates in federated learning on resource-constrained IoT microcontrollers. Implemented on ESP32-based edge devices, the framework combines light-weight block ciphers (SPECK, SIMON, and PRESENT), HMAC-SHA256 for integrity verification, and ECDH-HKDF for session-key establishment. The evaluation assessed latency, throughput, RAM/ROM footprint, and energy consumption. Results show that SPECK provides the lowest overhead (0.13 µs/byte, 8.68 MB/s, 138.3 mJ), SIMON offers intermediate performance (0.41 µs/byte, 1.96 MB/s, 184.9 mJ), and PRESENT incurs the highest computational cost (89.37 µs/byte, 0.011 MB/s, 446.2 mJ). In the CICIoT2023 federated intrusion detection evaluation, the secure model maintained stable convergence and achieved 85.43% accuracy after 20 rounds, remaining close to the centralized baseline. These findings demonstrate the practical feasibility of secure model-update exchange in FL on real IoT microcontrollers and provide hardware-grounded guidance for cipher selection under tight resource budgets.

This study presents the development of an Artificial Intelligence (AI)-based legal assistant using the Retrieval-Augmented Generation (RAG) architecture to provide legal assistance to citizens of the Republic of Kazakhstan. The proposed solution is designed to generate accurate, evidence-based responses to user queries using the regulatory legal acts of the Republic of Kazakhstan as the primary source of information. A legal corpus comprising 101,000 legislative documents and court decisions, with approximately 77 million tokens in Kazakh and Russian, was constructed to support the retrieval component of the system. To identify the most effective semantic retrieval method, three multilingual embedding models—Multilingual-E5-Large, BGE-M3, and KazEmbed-V5—were evaluated for vector search. The experimental results showed retrieval accuracies of 87.6%, 76.8%, and 83.3%, respectively. The GPT-5.4 and Llama-4-Scout-17B-16E-Instruct large language models were used to generate legal reasoning and responses based on documents retrieved through semantic search. The quality of the generated responses was evaluated using two complementary approaches. First, legal experts assessed the factual correctness and legal validity of the answers. Second, automatic evaluation was performed using word-level F1, BLEU, ROUGE, and BERTScore-F1 metrics. Among all evaluated configurations, GPT-5.4 combined with Multilingual-E5-Large achieved the highest overall accuracy (88.5%), whereas Llama-4-Scout-17B-16E-Instruct combined with KazEmbed-V5 achieved an accuracy of 83.6%. Based on the proposed architecture and the selected semantic retrieval and language models, an AI legal assistant was developed and integrated into the “Adal Azamat” legal services platform providing users in Kazakhstan with practical access to AI-assisted legal consultation.

Speech segmentation is the process of dividing speech signals into parts, which is an important aspect of speaker identification and speech recognition systems. This process improves the efficiency of the system by accurately detecting the beginning and end of speech. The use of voice activity detectors (VADs) plays an important role in segmentation, as they help to determine the boundaries between speech and silence. However, the most common errors in segmentation are false positives and false negatives, which negatively affect the overall accuracy of the system. In this regard, it is necessary to reduce errors through various approaches and methods. Measures such as reducing background noise, using deep learning models, and data augmentation can significantly improve the quality of segmentation. Using spectral analysis methods and features allows you to clearly distinguish between speech and background noise. The purpose of this study is to optimize the segmentation process and analyze the probability of errors, improve the efficiency of speech recognition systems. As a result, this work provides a basis for new research and development in the field of speech recognition. The article considers the problem of speech segmentation for speaker identification. The paper describes possible segmentation criteria – qualitative and quantitative characteristics of sound speech, such as speech delays and intonation, as well as their acoustic relationship. This allows a specialist to identify specific segment units (syllables, words, etc.), record their structure, and identify the main features.

Recent advancements in Legal Artificial Intelligence (LegalAI) have focused on automated legal case analysis and judgment prediction using Transformer-based language models, legal knowledge retrieval, structured case information and multimodal representations. Recent studies have also explored retrieval-augmented reasoning and multi-agent approaches to integrate case facts, applicable laws, similar cases, and contextual legal evidence. However, existing approaches often treat legal text, structured case information, external legal knowledge, and collaborative reasoning as separate components, limiting comprehensive context-aware judgment prediction. This work develops a knowledge-augmented multi-agent AI framework that integrates contextu-al legal text, structured case information, legal knowledge retrieval, and coordinated multi-agent reasoning to predict legal judgment outcomes as Approved or Rejected. The study uses a Legal Judgment Prediction Dataset containing 1,522 cases, including 1,000 Approved and 522 Rejected cases. The framework integrates Robustly Optimized Bidirectional Encoder Representations from Transformers (RoBERTa), Multimodal Feature Fusion (MFF), embedding-based retrieval with FAISS, and a Heterogeneous-Agent Proximal Policy Optimization (HAPPO)-based multi-agent system comprising Legal Retrieval, Similar Case, Case Analysis and Legal Reasoning Agents. Knowledge augmentation and context-aware reasoning are followed by Feature Tokenizer Transformer (FT-Transformer)-based outcome prediction. Experimental evaluation achieved 0.9876 accuracy, 0.9842 precision, 0.9918 recall, 0.9880 F1-score, and 0.9999 ROC-AUC, while re-trieval achieved 0.5843 Precision@5, 0.6371 Recall@5, and 0.7401 Mean Reciprocal Rank (MRR). Ablation analysis showed reduced performance after removing text embedding, multimodal fu-sion, knowledge retrieval, or HAPPO-based reasoning. Overall, the framework demonstrates an effective approach for context-aware legal judgment prediction and provides a foundation for validation on larger and more diverse legal datasets.