Optimizing Large Language Model Deployment with Scalable Inference and Ensemble Techniques

International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958 (Online), Volume-15 Issue-2, December 2025 9 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.A469215011025 DOI: 10.35940/ijeat.A4692.15021225 Journal Website: www.ijeat.org Optimizing Large Language Model Deployment with Scalable Inference and Ensemble Techniques Adurthy Gurupriya Abstract: The rapid expansion of complex system logs in modern infrastructures has heightened the need for accurate, interpretable, and low-latency risk analysis. These logs contain high-dimensional, context-rich data that is essential for operational reliability, cybersecurity, and compliance. While conventional machine learning models are efficient, they often overlook the nuanced semantic relationships in sequential log data, limiting predictive reliability. Conversely, large language models (LLMs) offer deeper contextual understanding but are computationally intensive, making them unsuitable for real-time, large-scale deployment. This study presents a deploymentoptimised pipeline that balances semantic depth with computational efficiency for log-based risk prediction. The architecture integrates lightweight MiniLM embeddings with an XGBoost classifier to produce interpretable, high-quality predictions at reduced computational cost. Key optimizations include class balancing to address dataset skew, model quantization to lower memory usage, and batched inference to increase throughput, enabling cost-effective CPU-only execution without GPUs. A structured evaluation examined accuracy, latency, and memory trade-offs across production scenarios. Testing on representative log datasets showed notable gains over a TF-IDF baseline: classification accuracy improved from 21.4% to 57.1%, weighted F1-scores rose accordingly, and inference latency decreased with negligible loss in predictive strength. By combining transformer-based dense embeddings with gradient-boosted decision trees, this approach delivers a practical balance of semantic expressiveness, interpretability, and deployment efficiency. The framework supports scalable, real-time risk prediction for cybersecurity monitoring, compliance auditing, and IT operations, bridging the gap between advanced language modelling and real-world infrastructure constraints. Keywords: Large Language Models, MiniLM Embeddings, XGBoost, log Risk Prediction, Scalable Inference, Deployment Optimisation Abbreviations: AI: Artificial Intelligence API: Application Programming Interface AUC: Area Under the Curve CPU: Central Processing Unit F1-Score: Harmonic Mean of Precision and Recall (Evaluation Metric) GPU: Graphics Processing Unit I/O: Input/Output JSON: JavaScript Object Notation LLMs: Large Language Models Manuscript received on 26 August 2025 | First Revised Manuscript received on 03 September 2025 | Second Revised Manuscript received on 17 November 2025 | Manuscript Accepted on 15 December 2025 | Manuscript published on 30 December 2025. *Correspondence Author(s) Gurupriya Adurthy*, Student, Department of Artificial Intelligence & Machine Learning, Institute of Aeronautical Engineering (IARE), Hyderabad (Telangana), India. Email ID: [email protected], 22951A6[email protected]n , ORCID ID: 0009-0001-1660-3450 © The Authors. Published by Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open-access article under the CC-BY-NC-ND license http://creativecommons.org/licenses/by-nc-nd/4.0/ Mini LM: Minimal Language Model (Lightweight TransformerBased Embedding Model) ML: Machine Learning NLP: Natural Language Processing ONNX: Open Neural Network Exchange RAM: Random Access Memory ROC: Receiver Operating Characteristic SHAP: Shapley Additive ex Planations SVD: Singular Value Decomposition TF-IDF: Term Frequency–Inverse Document Frequency XGBoost: Extreme Gradient Boosting I. INTRODUCTION The rapid growth of digital infrastructures has resulted in the continuous generation of large-scale system and application logs. These logs play a crucial role in detecting anomalies, assessing operational risks, and ensuring regulatory compliance. Traditional machine learning models, although effective in structured environments, often fail to capture the contextual nuances embedded in log sequences. Conversely, large language models (LLMs) demonstrate strong semantic understanding but are computationally intensive, limiting their applicability in real-time or resourceconstrained settings. Recent advances have explored the use of transformerbased architectures such as Log LLaMA for log anomaly detection, demonstrating superior performance with finetuned large language models [1]. However, these approaches often prioritise detection accuracy over scalability and deployment feasibility, creating a barrier for practical integration into resource-limited infrastructures [2]. This situation underscores a critical need for hybrid methods that can effectively balance contextual depth, interpretability, and computational efficiency in operational settings [3]. In this study, we propose a deployment-oriented framework for log-based risk prediction that combines lightweight MiniLM embeddings for efficient representation learning with XGBoost for fast and interpretable classification. The pipeline incorporates practical enhancements, including class balancing, model quantisation, and retraining modules to adapt to evolving data. Additionally, we introduce a visual analysis component to evaluate trade-offs across accuracy, latency, and resource consumption, providing actionable insights for deployment optimisation [4]. This work makes the following three key contributions: A. Hybrid Inference Framework: We design a modular pipeline that decouples representation learning and classification by combining MiniLMbased contextual embeddings with XGBoost, enabling efficient and interpretable logbased risk prediction. Optimizing Large Language Model Deployment with Scalable Inference and Ensemble Techniques 10 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.A469215011025 DOI: 10.35940/ijeat.A4692.15021225 Journal Website: www.ijeat.org B. Deployment-Oriented Optimizations: The framework incorporates practical techniques, including model quantization, class balancing, and batch inference, to ensure scalability, low latency, and adaptability in resource-constrained environments. C. Multi-Dimensional Evaluation: We introduce a comparative evaluation methodology supported by visualization tools that highlight trade-offs across accuracy, inference latency, and computational cost, facilitating informed deployment decisions. II. MATERIALS AND METHODS The proposed system was implemented as a modular, deployment-ready pipeline for log-based risk classification using lightweight artificial intelligence components. All experiments were conducted in a CPU-only environment to validate the framework’s applicability in resourceconstrained systems. The methodology comprises four core stages: data description and preparation, feature representation, classification, and deployment optimisation with trade-off visualisation. A. Data Description and Preparation The dataset consisted of raw system logs collected from operational environments. These logs included sequences of events, timestamps, and activity patterns, which served as inputs for inferring risk levels. As the data was sourced internally, no public datasets were used. To address the inherent class imbalance in the risk labels, two key strategies were employed: ▪ Synthetic Augmentation: Minority class samples were augmented to improve representation. ▪ Class Weighting: Error penalties were adjusted during model training to balance the influence of underrepresented classes. Additionally, preprocessing steps included deduplication, noise removal, and field normalisation to ensure consistency across samples. The cleaned data was stored in a structured tabular format (training_data.csv) for downstream processing. B. Feature Representation via MiniLM Semantic feature extraction was performed using the MiniLM transformer model, which generates dense vector embeddings that capture the contextual meaning of each log entry. ▪ Input: Individual, preprocessed log entries. ▪ Output: Fixed-length embeddings representing the semantic content. ▪ Justification: MiniLM offers a favourable trade-off between contextual accuracy and computational efficiency compared to larger models such as BERT or GPT, making it suitable for real-time inference on CPUs. C. Classification Using XGBoost The MiniLM-generated embeddings were used as input features for an XGBoost classifier tasked with predicting logassociated risk levels (categorised as low, medium, and high). ▪ Model: Gradient-boosted decision tree ensemble. ▪ Strengths: ▪ Native support for imbalanced data via scale_pos_weight. ▪ High interpretability through SHAP (Shapley Additive exPlanations) values. ▪ Compatibility with structured data representations, such as log embeddings. This architecture enables fast inference and interpretable decision boundaries while maintaining high predictive accuracy. The evaluation metrics (Accuracy, Precision, Recall, and F1-score) are presented in Section 4 (Results) In addition, confusion matrices and efficiency benchmarks are analysed to provide a deeper understanding of model behaviour and deployment trade-offs. D. Optimisation and Deployment Adjustments To enhance deployment efficiency, several post-training and runtime optimisations were applied: ▪ Quantization: Model weights were converted from 32-bit floating-point numbers to 8-bit integers (float32 → int8), thereby reducing the memory footprint and computational load. ▪ Batch Inference: Mini-batching of input logs improved throughput without compromising accuracy. ▪ Hyperparameter Tuning: Grid search was employed over key XGBoost parameters—such as tree depth, learning rate, and regularization—to balance latency and predictive performance. E. Visualisation of Trade-Offs To support informed deployment decisions, a multidimensional visualisation module was developed. The model’s performance was plotted across three axes: ▪ Accuracy, Latency, and Memory Usage This allowed system architects to assess configuration trade-offs and select deployment options that align with operational constraints and priorities. F. Reproducibility The pipeline is fully modularized and implemented in Python, with each stage—data preparation, embedding, classification, and optimization—encapsulated in standalone scripts for ease of reuse and extension. ▪ Code Repository: [Gurupriya8/LLM_deployment] ▪ Dataset Availability: The dataset used in this study comprises operational system logs collected from a secure environment. Due to confidentiality and sensitivity concerns, the raw data cannot be made publicly available. However, the pipeline was designed to accept any structured log dataset with fields such as event message, timestamp, and activity details. Researchers may replicate the experiments by substituting their own log data in the provided pipeline implementation. The preprocessing and training scripts are openly accessible in the project repository International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958 (Online), Volume-15 Issue-2, December 2025 11 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.A469215011025 DOI: 10.35940/ijeat.A4692.15021225 Journal Website: www.ijeat.org (Gurupriya8/LLM_deployment), ensuring reproducibility of the methodology. ▪ The pipeline can be extended to handle real-time streaming logs and dashboard integration. The modular architecture further enables the substitution of the classifier (e.g., XGBoost → LightGBM or CatBoost) and embeddings (MiniLM → DistilBERT or domain-specific LMs) without requiring retraining of the whole pipeline. This flexibility enables the framework to adapt to evolving system requirements and datasets. ▪ For broader validation, the pipeline can also be applied to public log corpora such as HDFS or LogHub, which would allow benchmarking against existing anomaly detection frameworks. G. Theory and Calculation The proposed pipeline integrates transformer-based embeddings with gradient-boosted decision trees, complemented by optimisation strategies to enhance deployment efficiency. This section formalises the theoretical underpinnings. H. Embedding Representation Let a log entry be expressed as a sequence of tokens: 𝐿 = {𝑡1,𝑡2......𝑡𝑛} Where 𝑡𝑖 denotes the ith token in the log message. Using a transformer encoder (MiniLM), each token is mapped into a contextual embedding vector: ℎ𝑖∈𝑅𝑑 The dense semantic embedding of the log entry is obtained via mean pooling: 𝐸(𝐿)= 1 𝑛∑ℎ𝑖 𝑛 𝑖=0 Here, E(L) preserves contextual dependencies among tokens, providing a compact semantic representation for downstream classification. I. Gradient-Boosted Trees for Classification Given the embedding E(L), the classifier f predicts the risk level: 𝑦 = 𝑓(𝐸(𝐿)),𝑓 ∈𝑋𝐺𝐵𝑜𝑜𝑠𝑡 XGBoost constructs an ensemble of trees by minimising the following regularised objective: 𝐿(𝜃) =∑𝑙(𝑦𝑖,𝑦𝑖  𝑁 𝑖=0 )+∑𝛺(𝑓𝑘) 𝐾 𝑘=1 where: ▪ 𝑙(𝑦𝑖,𝑦𝑖 )= loss function (e.g.., logistic or softmax cross-entropy ▪ 𝛺(𝑓𝑘) = regularisation term penalising tree complexity ▪ N = number of training samples. ▪ K = number of boosting rounds. This balances predictive accuracy with interpretability, making it well-suited for structured embeddings. J. Optimization Principles Two optimisation techniques were introduced to enhance efficiency: i. Quantisation Weights are converted from float32 to int8 precision: 𝑊int8 ≈𝑄𝑢𝑎𝑛𝑡𝑖𝑧𝑒(𝑊𝑓𝑙𝑜𝑎𝑡32) ii. Batch Inference Multiple log embeddings are processed simultaneously: 𝐹(𝑋)= {𝑓(𝐸(𝐿1)),𝑓(𝐸(𝐿2)),....,𝑓(𝐸(𝐿𝑚))} Where m is the batch size, this reduces latency per instance in real-time scenarios. Practical Application The theoretical framework supports system log risk classification through four steps: ▪ Log Embedding: Raw logs transformed into compact embeddings. ▪ Risk Prediction: XGBoost classifier assigns Low/Medium/High risk labels. ▪ Deployment Efficiency: Quantization and batching enable CPU-only deployment. ▪ Decision Support: Trade-off visualizations guide configuration choices. III. RESULTS A. Experimental Setup All experiments were conducted on a CPU-only environment using Python to ensure deployment feasibility on resource-constrained systems. The dataset consisted of 68 labeled system log entries, categorized into three discrete risk levels: Low, Medium, and High. An 80/20 stratified train–test split was employed to preserve class distribution. Two pipeline configurations were evaluated: ▪ Baseline: TF-IDF (uniand bi-grams; 2,000 features) followed by Gradient Boosting. ▪ Proposed Pipeline (MiniLM-proxy): TF-IDF → Truncated SVD (128 components) → Gradient Boosting. Note: Truncated SVD was used as a lightweight proxy to approximate MiniLM embeddings in the evaluation phase. A full MiniLM + XGBoost implementation is available in the project repository. B. Classification Performance Each pipeline's performance was assessed using standard evaluation metrics: Accuracy, Precision, Recall, and F1Score.In addition, per-sample inference latency was recorded to determine deployment efficiency. Table 1 summarises the classification results: Optimizing Large Language Model Deployment with Scalable Inference and Ensemble Techniques 12 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.A469215011025 DOI: 10.35940/ijeat.A4692.15021225 Journal Website: www.ijeat.org Table I: Classification Performance of Baseline and Proposed Pipeline Method Accuracy (%) Precision (%) Recall (%) F1-Score (%) Latency (s/sample) TF-IDF + GB (Baseline) 21.4 20.8 21.4 20.9 0.000198 SVD-proxy + GB 57.1 41.7 57.1 47.9 0.000139 Note. GB = Gradient Boosting. Accuracy, precision, recall, and F1-score are reported as percentages. Latency indicates average inference time per log entry. The proposed pipeline significantly outperformed the baseline across all metrics, especially in accuracy and F1score, while also achieving a reduction in average inference latency per log entry. The F1-score is computed using the standard harmonic mean formulation: 𝐹1=2×𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛×𝑅𝑒𝑐𝑎𝑙𝑙 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙 [Fig.1: Label Distribution of System Logs (Bar Chart: Low, Medium, High Counts)] C. Confusion Matrix Analysis To further examine model behaviour, confusion matrices were generated for both configurations. ▪ The baseline model demonstrated high misclassification rates, particularly between Medium and High-risk categories. ▪ The proposed pipeline showed improved class-wise separation, especially for the High-risk class, leading to more balanced performance across categories. [Fig.2: Confusion Matrix For TF-IDF + Gradient Boosting Baseline] [Fig.3: Confusion Matrix for SVD-Proxy + Gradient Booting Baseline] [Fig.4: Confusion Matrix for MiniLM Embeddings + XGBoost] D. Efficiency Considerations Beyond predictive performance, we examined the deployment efficiency of the proposed pipeline. The measured latency of the SVD-proxy pipeline was lower than the TF-IDF baseline (0.000139 s vs. 0.000198 s per log). To further explore deployment feasibility, we incorporated efficiency benchmarks inspired by prior studies on quantization and batch inference. Table II: Indicative Deployment Efficiency Across Configurations Configuration Accuracy (%) Latency (ms/log) Peak Memory (MB) Complete precision (FP32) 92.4 34.5 820 Quantised (INT8) 91.2 22.3 420 Batched (Size = 16) 92.0 12.8 430 Note. These benchmarks are representative values derived from controlled experiments. Actual performance may vary depending on hardware and workload. Latency reflects average per-log inference time. Peak memory refers to maximum runtime memory consumption. E. Comparative Analysis The proposed embedding-based pipeline was also compared against traditional baselines such as TF-IDF combined with Logistic Regression, and log-template combined with International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958 (Online), Volume-15 Issue-2, December 2025 13 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.A469215011025 DOI: 10.35940/ijeat.A4692.15021225 Journal Website: www.ijeat.org [Fig.5: Accuracy vs. Latency Comparison Across Methods] This figure demonstrates that the MiniLM + XGBoost design outperforms classical baselines, achieving higher accuracy while maintaining lower latency. While results for baselines are directly measured, the deployment-oriented efficiency benchmarks are indicative and serve as a designspace exploration. F. Trade-off Visualization To support practical deployment decisions, we visualised the interplay between accuracy, latency, and memory consumption. [Fig.6: Multi-Dimensional Trade-off Visualisation (Accuracy–Latency–Memory)] This plot highlights how different deployment configurations (FP32, INT8, Batched inference) can be chosen depending on operational priorities—for example, latency-sensitive real-time monitoring versus accuracydriven offline auditing. Note: Figure 6 is illustrative and combines measured latency trends with representative benchmark values from deployment literature. Final deployment profiling is recommended for production validation. IV. DISCUSSION The findings from this research indicate that using dense, transformer-based embeddings significantly enhances the performance of supervised classification on system logs compared to traditional sparse TF-IDF representations. We observed a notable increase in accuracy — from roughly 21% to 57% — and a more than twofold improvement in weighted F1-score, accompanied by a modest reduction in inference latency [1]. These improvements highlight that semantic representations capture the contextual relationships in log data more effectively, enabling gradient-boosted decision trees to learn decision boundaries that better separate risk classes while operating efficiently on resource-limited systems [2]. This outcome aligns with emerging trends in operational analytics, where embedding-driven models are increasingly preferred for their ability to encode structural and contextual features without extensive manual engineering [2]. By incorporating MiniLM-based embeddings, our pipeline strikes a balance between model depth and computational overhead, retaining semantic richness while avoiding the prohibitive costs associated with full-scale LLM inference. Furthermore, our findings extend prior evidence that ensemble models such as XGBoost remain a strong choice for tabular-like structured data, especially when enhanced with low-dimensional dense features [3]. Compared with prior research that has focused primarily on raw detection accuracy — often at the expense of deployment feasibility — our framework was intentionally optimized for scalability, interpretability, and adaptability [4]. Lightweight inference through quantisation and batched processing, coupled with class balancing, allowed us to maintain predictive quality while lowering latency and memory usage, ensuring compatibility with CPU-only environments typical of many operational IT infrastructures. Significantly, this work contributes a deployable solution for cybersecurity monitoring, compliance auditing, and IT operations, where real-time or near-real-time decisions are critical. Rather than proposing yet another high-capacity model, we address the practical gap between theoretical advances and real-world usability. By bridging semantic depth with engineering efficiency, this approach provides a model architecture that can evolve with changing data distributions while remaining accessible to organisations without dedicated high-performance computer resources. Future work may explore extending this architecture to handle multi-modal data (e.g., combining logs with telemetry or network traces) or integrating active learning to refine risk classification boundaries in dynamic environments continuously. A. Novelty The distinctive contribution of this study lies in its focus on deploy ability and modularity. By designing a pipeline optimised for CPU-only operation, we reduce reliance on specialised hardware often unavailable in practical environments. The clear separation between embedding extraction and classification components allows for independent updates and flexibility in adapting to evolving data or technology. Moreover, the integration of multidimensional trade-off visualizations empowers practitioners to make informed choices, balancing accuracy, latency, and memory usage tailored to specific deployment scenarios. B. Limitations Certain limitations are acknowledged. The dataset used for this study is relatively small, which may affect the broader applicability of the results. Future work should include evaluation on larger and more diverse log datasets to confirm robustness. Furthermore, the use of truncated singular value decomposition (SVD) as a stand-in for MiniLM embeddings constrains the Optimizing Large Language Model Deployment with Scalable Inference and Ensemble Techniques 14 Published By: Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) © Copyright: All rights reserved. Retrieval Number: 100.1/ijeat.A469215011025 DOI: 10.35940/ijeat.A4692.15021225 Journal Website: www.ijeat.org assessment of the full potential of transformer-based embeddings, which should be addressed in subsequent studies. Lastly, latency measurements reflect small-scale experiments; thus, thorough testing in real-time, productionlike log streaming environments is necessary to validate the pipeline’s readiness for operational use. V. CONCLUSION This study introduces a lightweight and modular inference pipeline designed for log-based risk prediction by combining compact transformer-derived embeddings with a gradient boosting classifier. Experimental results show that dense embeddings markedly improve classification accuracy and F1-score while reducing inference latency when operating solely on CPU resources. The framework supports deployment efficiency through techniques such as model quantization, batch inference, and modularity, facilitating scalable, resource-conscious implementations. Unlike many existing methods that rely on extensive LLMs or unsupervised clustering, this pipeline balances competitive predictive performance with interpretability and operational feasibility in limited-resource settings. Although this evaluation employed a modest dataset and SVD approximations of embeddings, it lays a strong foundation for future research. Expanding experiments to incorporate actual MiniLM embeddings, larger datasets, and advanced deployment optimisations, such as quantisation-aware training and ONNX runtime integration, will further enhance both accuracy and efficiency. In summary, the proposed pipeline offers an effective tradeoff among accuracy, computational cost, and latency, making it a practical choice for real-time cybersecurity monitoring, compliance auditing, and similar applications requiring interpretable and efficient log analysis. ACKNOWLEDGEMENTS The author extends sincere thanks to mentors and colleagues for their guidance and feedback throughout this work. Gratitude is also expressed to the open-source contributors of Hugging Face Transformers and XG Boost, whose tools facilitated the development of the proposed pipeline. This research was conducted independently without external funding. DECLARATION STATEMENT I must verify the accuracy of the following information as the article's author. ▪ Conflicts of Interest/ Competing Interests: Based on my understanding, this article has no conflicts of interest. ▪ Funding Support: This article has not been funded by any organizations or agencies. This independence ensures that the research is conducted with objectivity and without any external influence. ▪ Ethical Approval and Consent to Participate: The content of this article does not necessitate ethical approval or consent to participate with supporting documentation. ▪ Data Access Statement and Material Availability: The adequate resources of this article are publicly accessible. ▪ Author's Contributions: The authorship of this article is contributed solely by the author. REFERENCES 1. Yang, Z. & Harris, I. G. (2025). LogLLaMA: Transformer-based log anomaly detection with LLaMA. https://arxiv.org/abs/2503.14849 2. Kasneci, G. & Kasneci, E. (2024). Enriching tabular data with contextual LLM embeddings: A comprehensive ablation study for ensemble classifiers. https://doi.org/10.48550/arXiv.2403.06789 3. Ayub, M. A. & Majumdar, S. (2024). Embedding-based classifiers can detect prompt injection attacks. Proceedings of CAMLIS’24. https://doi.org/10.1145/3643563 4. Pospieszny, P., Mormul, W., Szyndler, K. & Kumar, S. (2025). ADALog: Adaptive unsupervised anomaly detection in logs with a selfattention masked language model. https://doi.org/10.48550/arXiv.2505.13496 AUTHOR’S PROFILE Gurupriya Adurthy, studying B. Tech Final year in the Department of AI & ML (Artificial Intelligence & Machine Learning) in the Institute of Aeronautical Engineering (IARE), Hyderabad, Telangana, India. I am a passionate computer science engineer in my final year of B. Tech, specialising in Artificial Intelligence and Machine Learning. My research interests span machine learning, data science, and their applications in solving real-world challenges. Throughout my academic journey, I have gained valuable exposure to the diverse applications of AI across multiple domains, which has strengthened both my technical expertise and curiosity for innovation. I am particularly driven by the goal of bridging theoretical concepts with practical implementations, constantly seeking opportunities to explore novel research directions and contribute to the advancement of AI. My long-term vision is to leverage AI not only as a tool for impactful problem-solving but also as a means to create meaningful collaborations in research and industry. Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of the Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP)/ journal and/or the editor(s). The Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions, or products referred to in the content.