Continuous Integration/Deployment (CI/CD) pipelines are critical for integrating developer changes and maintaining high-quality software deployments. The increasing frequency of commits and deployments places significant demands on CI/CD systems, requiring improved speed and efficiency. While numerous tools and techniques have been proposed to increase the velocity of CI/CD pipelines, there is a notable gap in architectural guidance for developers on key design decisions and best practices. To address this, we conducted a grey literature review using Straussian Grounded Theory to develop a UML-based model to guide software architects and developers in their decision-making. Our research focuses on identifying architectural design decisions (ADDs) and best practices as decision options that improve the speed and efficiency of CI/CD pipelines. The study analyses 38 sources, building a formal model comprising 6 ADDs and 30 best practices. This work contributes a structured, architecturally guided approach to optimizing CI/CD systems.

As artificial intelligence continues to advance, Reinforcement Learning (RL) has established itself as a core approach for developing intelligent agents that make decisions over time. As RL systems grow in complexity, the need for standardized training practices becomes critical. This paper introduces a rule-based assessment approach to enforce best practices in RL training. We define a comprehensive set of architectural rules focused on RL pipeline practices, models versioning, multi-agents deployment and managing models in inference. Our methodology integrates Large Language Models (LLMs) and custom-based code detectors to ensure compliance with these best practices across diverse RL systems. We developed a ML pipeline insights service to automatically validate RL training practices directly from the source code. We validate our approach by applying it in a large-scale industrial case study and sixteen open-source case studies. Our evaluation showed that custom-based detectors achieved near-perfect precision and recall ( $$ F_1 \approx 0.98 $$ ), while LLM-based detectors provided scalable validation with moderate $$ F_1 $$ scores (0.67–0.71), demonstrating the hybrid approach’s strength in balancing accuracy and automation. The results demonstrate our tool’s accuracy in identifying and enforcing best practices with high precision and recall rates, highlighting its practical applicability and automation feasibility.