Leandro Augusto Silva

Asset maintenance decision-making in complex systems requires the integration of heterogeneous operational information, formal optimization models, and reliable decision-support mechanisms. However, a significant portion of maintenance knowledge is expressed in natural language, creating a gap between real-world problem descriptions and the structured inputs required by Operations Research (OR) models. This paper presents LLM-OR Optimization (LLMOR-Opt), a reproducible method that combines Large Language Models (LLMs) and OR to support the generation of optimal asset maintenance decisions. The method uses LLMs to interpret maintenance requirements, formulate optimization problems, and create a structured intermediate representation in JavaScript Object Notation (JSON). Human validation is incorporated before generating executable Python code, while computational validation is performed before solver execution. OR solvers are then responsible for obtaining feasible or optimal solutions, which are subsequently translated into natural language explanations to support maintenance planners. By separating semantic interpretation, mathematical formulation, computational implementation, optimization, and explanation, LLMOR-Opt improves traceability and provides a structured approach for integrating generative AI with formal optimization methods in asset maintenance decision-making.• Integrates LLM-based interpretation with formal OR optimization.• Introduces human and computational validation checkpoints.• Provides traceable and explainable maintenance decision support.

Context: Modularization of monolithic systems directly influences comprehension, maintenance, and architectural restructuring activities. Automatic identification of candidate modules from structural dependencies is a recurring strategy, but empirical evidence comparing classical community detection heuristics with embedding-based methods under equivalent experimental conditions remains limited.Objectives: This study compares modularity maximization heuristics (Louvain and Greedy Modularity) and embedding-based methods (Node2Vec and Graph Autoencoders followed by k-means) applied to the structural modularization of monolithic Java systems, evaluating topological quality, architectural coherence, and computational scalability.Methods: The methods are evaluated on undirected and weighted graphs extracted via static analysis from eight open-source Java systems, ranging from 43 to 29,695 classes. The evaluation combines weighted modularity, object-oriented metrics (LCOM4, CBO, RFC, NOM, NOA) aggregated by community, partition stability (NMI, ARI), and execution time.Results: Classical heuristics achieved modularity values up to 0.91 across all systems, consistently outperforming embedding-based methods (at most 0.73 for Node2Vec and 0.41 for GAE), while requiring 10 to 60 times less computational cost. Graph autoencoders produced stable but structurally undifferentiated partitions on sparse graphs without node attributes.Conclusion: For structural modularization based exclusively on static dependencies, classical heuristics remain highly competitive. Embedding-based methods do not present clear advantages in either partition quality or computational cost unless enriched with semantic or attribute information.

JASTG (Java Abstract Syntax Tree Structural Graph) is an open source tool for deterministic extraction of structural dependency graphs from Java systems using AST-based static analysis. It identifies explicit typed dependencies, including inheritance, interface implementation, field types, method parameters and instantiations, and exports weighted graphs in GraphML format. Object-oriented metrics such as LCOM4, CBO, RFC, NOM and NOA are computed per node. The tool operates directly on source code without build system integration, ensuring environment-independent and reproducible results. A versioned benchmark dataset comprising graphs and metrics from 21 Java systems is publicly available on Zenodo, enabling independent validation and reuse in structural modularity research.

Project-Based Learning (PBL) in software engineering produces heterogeneous evidence across repositories, task boards, classroom activities, and assessments, making timely supervision difficult. This paper presents a two-stage method to model PBL digital twins with the five RM-ODP viewpoints, aligned with ISO/IEC 10746 and ISO/IEC/IEEE 42010 \cite {iso10746, iso42010}. Stage~A models the as-is PBL ecosystem with SADT/IDEF0 and identifies native event points and requirement gaps. Stage~B adds a Digital Twin layer that consumes those events, applies data contracts, correlates patterns through Complex Event Processing (CEP), and delivers near real-time feedback without replacing existing processes. The method structures: goals and governance (enterprise), semantic contracts and traceability (information), event services and CEP rules (computational), distributed pipeline and immutable snapshots (engineering), and implementation choices (technology). An institutional case instantiation demonstrates end-to-end applicability from process modeling to operational supervision signals. The result is a reusable procedure for closed-loop supervision, auditable evidence integration, and longitudinal learning analytics in diverse PBL settings.

PBLRepositoryMetrics is a teacher-support system for Project-Based Learning (PBL) environments. It captures repository events from GitLab and evaluates them against a Software Configuration Management (SCM) baseline defined by a metaproject (Sprint Descriptor), using Complex Event Processing (CEP) to compute daily conformance distances. A four-phase pipeline performs static verification, dynamic CEP pattern detection, and LLM-based semantic evaluation, producing per-deliverable diagnostic reports for evaluators. The baseline evolves through professor assessment, preserving the professor as the final interpreter.

This article proposes an innovative conceptual model for Machine Learning Operations (MLOps) pipelines, aiming to overcome the current challenges concerning the entire lifecycle of machine learning models and to meet the growing demands of both Academia and Industry. Based on a hybrid research approach, combining scientific works and insights from professionals in the field, this proposed MLOps pipeline model integrates advanced automation, robust governance, intelligent data and model management, and explainable monitoring. We explore the convergence between theory and practice, identifying gaps and proposing an approach that promotes the scalability, reproducibility, and reliability of ML systems in complex and dynamic production environments. A state-of-the-art conceptual model for MLOps pipelines was proposed, based on a rigorous analysis of the literature and valuable insights from professional practice. The model addresses the critical challenges of automation, data and model management, monitoring, governance, and usability, aligning research ambitions with operational needs. The results from applying the MLOps architecture demonstrated measurable efficiency with a perceived improvement in the scalability, reproducibility, and reliability of ML systems. Positive outcomes were observed in relation to the deployment time of Machine Learning models, which was reduced from approximately 6 months to a range of 3 to 5 days, depending on the team’s maturity and the application’s purpose. An increase in productivity and operational standardization was also noted, accompanied by gains in scalability and efficiency, evidenced by the elimination of the model deployment queue, the migration of over 3,200 users to the new environment, and the publication of more than 100 Data Science models in the first few months of the new environment’s operation. Additionally, the transition to a cloud infrastructure provided cost and financial resource optimization compared to the previous on-premises solution, and an enhancement of governance and security through the execution of standardized pipelines.

Monolithic architecture is a software design which brings significant difficulties to system developers when it comes to maintenance or expanding the scope of a project. On the other hand, a modular project consists of several similar entities, or modules, which are the object of similar functions or processes that, applied repeatedly, have well-defined classes and smaller modules to work, bringing benefits such as reduced project development time and increased productivity for the system developers. This work proposes the use of complex networks through the NetworkX library in Python, using modularity detection algorithms for the static analysis of Java code. The goal is to discover modules by analyzing dependencies between classes, indicating the best way to identify code clusters to be treated as modules automatically. The outcomes of applying the Greedy Modularity, Louvain, K-Clique, and Girvan Newman algorithms to two open-source projects will be presented. A comparative analysis of these results will be illustrated using generated graphs and a distribution map, emphasizing the number of communities identified by each algorithm.

Context: Machine learning has become an essential tool for addressing complex problems in information systems, encompassing industrial, commercial, and residential applications. Problem: Machine learning systems without frequent retraining are prone to data and concept drift, compromising predictive accuracy. This issue is particularly critical in scenarios where retraining is infeasible due to high computational costs or data unavailability. Solution: This study evaluates the performance of drift detection methods in discrete time series with controlled changes in mean and standard deviation using synthetic Gaussian signals. IS Theory: The General Systems Theory underpins the study by emphasizing how the interplay between drift detection and adaptive systems contributes to maintaining stability and efficiency in dynamic environments. Method: Experiments were conducted with variations in mean, standard deviation, and both parameters simultaneously in order to obtain qualitative patterns of drift detectors behaviors. The detectors ADWIN, KSWIN, and Page-Hinkley were tested under this scenario. Summary of Results: The findings reveal that ADWIN and Page-Hinkley exhibited greater precision and robustness, while KSWIN showed excessive sensitivity, leading to a high number of false positives. Contributions to the IS Field: This research offers a comprehensive analysis of drift detectors’ performance, specifically in scenarios involving changes in mean and standard deviation, providing useful reference for designing resilient machine learning-based forecasting systems. Impacts on the IS Field: The study advances the development of information systems that can adapt to dynamic data environments characterized by shifts in mean and standard deviation, with direct applications in industrial contexts and energy management.