Engineering Software Research and Development (United States)

The article examines Human-in-the-Loop (HITL) architectures for automated decision-making systems deployed in enterprise operations and regulated domains. The topic’s relevance follows from the rapid adoption of high-capacity models alongside stricter requirements for accountability, traceability, explainability, and risk control. The paper’s novelty lies in formalizing a taxonomy of intervention modes and linking engineering choices to operational metrics rather than to model accuracy alone. The study identifies four recurring intervention patterns—pre-emptive review, confidence-based routing, asynchronous audit, and exception handling—and specifies their placement within the decision pipeline. The analytical basis relies on a comparative synthesis of documented production deployments in finance, healthcare, and corporate operations, focusing on throughput, decision quality, latency, and per-case processing cost. The results indicate a non-linear trade-off between automation rate and decision quality and show that optimal thresholding depends on risk asymmetry and governance constraints. Practical recommendations address uncertainty calibration, reviewer interface design, and closed-loop feedback capture for continuous improvement. The overall objective is to provide a deployment-oriented framework for selecting HITL patterns and tuning escalation thresholds in high-stakes settings.

Modernizing large-scale enterprise systems presents significant challenges due to tightly coupled legacy architectures, complex deployment dependencies, and high operational risks. The transition from monolithic applications to cloud-native microservices orchestrated by Kubernetes further amplifies deployment and scalability concerns, particularly in environments managing thousands of clusters. Traditional Kubernetes management platforms face inherent scalability limitations, creating bottlenecks in deployment governance, observability, and reliability. This paper presents a scalable continuous integration and continuous deployment (CI/CD) architecture leveraging multiple standalone Fleet controllers integrated with HAProxy for centralized routing and cluster management. By adopting a GitOps-driven deployment model using Fleet, combined with deterministic cluster-to-controller assignment and label-based rollout strategies, the proposed architecture efficiently manages over 10,000 Kubernetes clusters. HAProxy enables seamless access to distributed Fleet controllers through path-based routing, significantly reducing operational complexity and infrastructure overhead.The architecture incorporates automated fleet agent registration using SUSE Manager (MLM), eliminating manual intervention and ensuring balanced controller utilization. Performance optimizations—including etcd quota tuning and HAProxy connection scaling—address real-world operational constraints encountered during large-scale deployments. Experimental results demonstrate improved scalability, reduced deployment latency, and enhanced reliability. This work provides a practical reference architecture for enterprises seeking to implement resilient, large-scale Kubernetes CI/CD systems.

The dispersion coefficient of the constant phase element (CPE) is typically treated as an empirical fitting parameter in the analysis of impedance spectroscopy data, with no clear physical meaning. Here, we present an energy-based interpretation for this coefficient by linking it to the ratio of the dissipated or stored energy in the CPE relative to that supplied by the input source. Using the RC network equivalency of a CPE, we decompose the total input energy into a contribution stored in the capacitive modes and another dissipated in the resistive modes. Analytical expressions of these energy quantities are derived for three test examples: (i) a constant voltage, (ii) a voltage ramp, and (iii) a quadratic input of the form v(t) = λt2. In all cases, we found that the ratios of any two of these energy quantities reduce to pure functions of the dispersion coefficient of the CPE, independent of excitation amplitude or material parameters. This result provides a new perspective on the CPE’s dispersion coefficient from a thermodynamic/energetic basis, with direct implications for supercapacitor characterization, battery modeling, as well as for the analysis of other electrochemical devices and systems exhibiting the CPE behavior.

This paper proposes Edge-FIT (Federated Instruction Tuning on the Edge), a scalable framework for Federated Instruction Tuning (FIT) of Large Language Models (LLMs). Traditional Federated Learning (TFL) methods, like FedAvg, fail when confronted with the massive parameter size of LLMs [3], [6]. Our Edge-FIT framework combines federated learning with 4-bit Quantized Low-Rank Adaptation (QLORA), mitigating the core issues of communication and computational overhead. We demonstrate this by filtering the general-purpose Databricks Dolly 15k dataset for the IoT domain. Experimental results show the Edge-FIT-tuned Llama 2 (7B) achieves an F1-Score of 0.89. We also demonstrate a viable trade-off using the 3.8B Phi-3-mini model, validating Edge-FIT as a scalable framework for decentralized LLM deployment on home compute gateways.

The Front Door Request Tool is an innovative breakthrough in the release management of enterprises thathas been able to solve the long-running problems of the manual, time-consuming process of changerequests that have bedeviled the software development organizations. This AI-powered web app, developed by Raghu on behalf

With the growing dependence on APIs in contemporary software systems, API security has become a pressing challenge, especially against automated AI-based attacks and zero-day threats. Signature-based and rule-based security solutions are not adequate to identify well-structured pattern-based sophisticated reconnaissance or probing attacks. This study suggests an AI-based Security Platform (ASP) that learns dynamic patterns of valid API requests from actual application traffic. The system uses hierarchical semantic analysis to vectorize and flatten structured API payloads and then extract features. A self-attention mechanism augmented Recurrent Neural Network (RNN) captures the structural and temporal patterns in the request sequences. Anomaly detection policy classifies traffic in real time and only permits behavior conforming to the learned patterns while rate-limiting or blocking exploratory calls deemed anomalous. This strategy avoids attackers mapping system behavior and lowers the attack surface, with an added benefit of zero-day resilience. The new method makes large gains in accuracy and false-positive rate over current models.

With the rise of artificial intelligence (AI) in the ecosystem of contemporary cyber threat intelligence (CTI) platforms, the issue of AI-driven decision interpretability and transparency has become increasingly common. Although there is an increased ability in machine learning models to detect the complex and evolving cyber threats, this usually prevents human trust and restricts perceived actionable insights because they are black boxed and have a challenge of accountability. This paper discusses the use of Explainable Artificial Intelligence (XAI, including SHAP (SHapley Additive exPlanations), LIME (Local interpretable Model-Agnostic Explanations), and attention-based visualizations, in order to achieve better interpretability in CTI systems. We provide a case study illustrating that incorporation of XAI to threat detection pipelines enhances the analysis comprehension by analysts, saves investigation time, and enables informed decision-making within Security Operations Centers (SOCs). The evidence provided by our research indicates that XAI does not only make complex AI models understandable to human analysts yet can create a collaborative and transparent cybersecurity ecosystem. The study offers a viable system to implement XAI-enriched CTI that would enable more responsible and successful AI-empowered cybersecurity.

This study takes a mixed-methods approach to explore the impact of AIOps on IT Service Management process. Modern IT operations face unprecedented challenges in managing increasingly complex, distributed systems. This paper examines how Artificial Intelligence for IT Operations (AIOps) transforms traditional IT Service Management (ITSM) from reactive firefighting to proactive, intelligence-driven operations. This comprehensive framework demonstrates how organizations can implement AIOps across the Incident and Change management lifecycle to achieve quantifiable improvements in key performance metrics. Additionally, AIOps enables contextual enrichment of incidents, automated remediation workflows, and data-driven change risk assessment. This paper presents a maturity model for AIOps adoption, identifying critical success factors and implementation strategies that enable organizations to realize maximum ROI. Through analysis of implementation case studies across diverse industries, we document significant enhancements in operational efficiency, including average reductions of 73% in detection time and 62% in resolution time. Our findings illustrate how AIOps systematically addresses the fundamental challenges of modern IT environments by creating resilient operations that can anticipate and prevent disruptions before they impact users

One of the most important aspects of risk management for financial institutions is assessing credit risk. Credit scoring models are important tools for evaluating loan applications because they provide a systematic way to assess credit worthiness. While traditional statistical models have been widely used, artificial intelligence (AI) has emerged as a more efficient alternative due to its ability to process large datasets and improve predictive accuracy. The growing reliance on AI-powered models has transformed lending practices, improving decision-making, reducing default risks, and enhancing financial stability. The focus of this research is on exploring AI-based credit scoring models and their impact on financial institutions. Traditional credit scoring methods often lack accuracy and efficiency, leading to increased risks and losses. AI methods like machine learning and deep learning offer a more reliable method, analyzing huge amounts of data and spot patterns that people are not aware of. Gaining insight into how AI affects credit scoring helps with risk management, loan selection, and financial inclusion. Other options for A1, A2, A3, A4, and A5. Income level, credit score, existing debt, and recent credit inquiries are all part of the assessment. The results showed that A3 ranked lowest and A4 ranked best. A1 has the highest value for The Role of AI in Enhancing Credit Scoring Models for Better Lending according to the TOPSIS Method approach.