Abstract
Ensuring operational consistency and reproducibility in predictive modeling represents a critical challenge when transitioning models into production environments. To address this issue, this study proposes a cloud-based MLOps architecture designed to optimize portability and reproducibility throughout the predictive model lifecycle. The proposed approach decouples a deterministic ingestion pipeline based on heuristic rules, focused on data cleansing, temporal validation, and metadata normalization, from the inference engines and persistence layer. Kubernetes-based orchestration and Terraform provisioning through Infrastructure as Code (IaC) were implemented to guarantee functional equivalence between Microsoft Azure cloud services and on-premises environments. Additionally, the architecture incorporates an interpretability layer supported by Large Language Models (LLMs), capable of transforming statistical metrics into natural language narratives to improve analytical understanding. The architecture was validated through a case study using metadata from Ecuadorian scientific publications, demonstrating reduced deployment latency and operational consistency across environments. Empirical validation on 87,073 records spanning 2015–2023 showed that a multi-model competitive framework comprising Poisson GLM, Ridge, Lasso, Gradient Boosting, Random Forest, and ETS (Holt-Winters) selected ETS as the optimal model under the temporal cross-validation criterion (mean validation MAPE = 18%, composite score Scomp=4.33), while the final point-error evaluation for the 2026 forecast horizon yielded a MAPE = 11.2%. The selection process employed an expanding-window protocol (K = 3 folds) and conformal prediction intervals. The ETL pipeline completed normalization of 16,599 unique journal categories across 87,073 records in a total processing time of 27.91 seconds, confirming a throughput of approximately 3,120 records/second. The findings indicate that the proposed architecture minimizes configuration drift, facilitates scalability, and improves deployment efficiency in predictive analytics systems.

This publication has 36 references indexed in Scilit: