The microservice granularity directly affects the quality attributes and usage of computational resources of the system, determining optimal microservice granularity is an open research topic. Microservices granularity is defined by the number of operations exposed by the microservice, the number of microservices that compose the whole application, and its complexity and dependencies. This paper describes “Microservice Backlog (MB)”, a semiautomatic model for defining and evaluating the granularity of microservice-based applications; MB uses genetic programming technique to calculate at design time the granularity of each microservice from the user stories in the “product backlog” or release planning; the genetic algorithm combined coupling, cohesion, granularity, semantic similarity, and complexity metrics to define the number of microservices, and the user stories associated with each microservice. MB decomposes the candidate microservices, allowing to analyze graphically the size of each microservice, as well as its complexity, dependencies, coupling, cohesion metrics, and the number of calls or requests between microservices. The resulting decomposition (number of microservices and their granularity) performed by MB shows less coupling, higher cohesion, less complexity, fewer user stories associated with each microservice, and fewer calls among microservices. MB was validated against three existing methods, using two state-of-the-art applications (Cargo Tracking and JPet-Store), and one real-life application (Foristom Conferences). The development team and/or architect can use metrics to identify the critical points of the system and determine at design time how the microservice-based application will be implemented.

In agile software development processes, user stories (US) had been used to specify application functionalities from the users’ perspective. For intelligent applications leveraging artificial intelligence (AI), the Smart Product Backlog (SPB) has included both AI-implementable and non-AI functionalities. This paper had proposed a model employing supervised machine learning techniques to classify US descriptions based on their technical feasibility for AI implementation. This model had aimed to assist in constructing smart product backlogs (SPB). Classifying US in agile development, particularly with AI, had been a labor-intensive process demanding significant time and expert involvement. Additionally, the lack of a dedicated dataset for this task had limited the applicability of automated methods. This study addressed this challenge by having experts classify the Mendeley US dataset using binary classification (AI and non-AI). The proposal had involved developing an automatic classification model to process US descriptions and distinguish those suitable for AI implementation. This model had leveraged advanced text processing techniques to refine the textual features of the US. Additionally, it had employed three binary classification techniques: logistic regression, k-nearest neighbors (k-NN), and support vector machines (SVM). The model’s performance had been evaluated using metrics such as accuracy, loss (Log-Likelihood Loss), precision, recall, F1 score, area under the ROC curve, and specificity to identify the best-performing technique in binary classification. Logistic regression and SVM models had demonstrated high accuracy, with scores of 0.748 and 0.740, respectively. These results had highlighted the potential of an automated tool for recommending US feasible for AI development, thereby supporting decision-making in agile software projects.

In agile software development, defining functional requirements is essential for effective planning and implementation. This work proposes an artificial intelligence (AI)-based model to automatically classify user stories according to their feasibility for AI implementation, enabling the creation of an intelligent product backlog. A feedforward neural network (FNN) was implemented and evaluated using data from 22 software projects. The model's performance was compared against traditional approaches such as logistic regression and support vector machines (SVM), as well as large language models like Gemini and ChatGPT. The FNN outperformed all other models, achieving an accuracy of 80.3%. Automating this task helps optimize backlog management, reduce operational costs, and improve overall system efficiency, contributing to more agile, intelligent, and adaptive software development processes.

En Colombia, las pequeñas y medianas empresas de impresión 3D afrontan baja visibilidad, comisiones elevadas y riesgos de fraude al operar en plataformas de comercio electrónico generalistas. Este artículo presenta 3DFanatics, un prototipo de plataforma inteligente diseñado para optimizar la experiencia de usuario y reforzar la seguridad mediante recomendaciones personalizadas, atención conversacional automatizada y verificación biométrica. Se adoptó la metodología ágil Scrum en ocho sprints para iterar rápidamente sobre diseño e implementación. La arquitectura se desplegó en Amazon Web Services con CDK. AWS Cognito gestionó las identidades; API Gateway y funciones Lambda se encargaron de la lógica de negocio; RDS y S3 almacenaron los datos; y el frontend en React JS se desplegó en Vercel. La capa de recomendación se implementó con AWS Personalize, entrenada con un conjunto sintético de 25 usuarios, 16 productos y entre 1 000 y 10 000 interacciones, este sistema se evaluó mediante métricas de ganancia acumulada descontada normalizada (NDCG), precisión, rango recíproco medio (MRR) y cobertura, alcanzando una precisión de 0,75 al generar cinco sugerencias. El chatbot desarrollado con AWS Lex permitió gestionar operaciones CRUD de inventario y pedidos, aunque mostró limitaciones en consultas no previstas. Para la verificación de identidad se integró AWS Rekognition con prueba de vivacidad con un umbral del 80 % y comparación facial con un umbral del 75 %. Además, AWS Textract en el sistema de verificación extrajo datos de las cédulas de ciudadanía con un umbral de confianza del 80 %. El proceso de validación incluyó pruebas unitarias con Jest y Pytest, pruebas de integración y pruebas end-to-end con Cypress, así como sesiones de usabilidad think-aloud con cinco emprendedores del sector. El análisis de costos correspondiente al período de desarrollo entre octubre del 2024 y marzo del 2025 reveló un gasto acumulado inferior a cincuenta dólares estadounidenses, lo que demuestra la viabilidad económica del prototipo. En conclusión, 3DFanatics valida su factibilidad técnica y económica como solución de comercio electrónico especializada en impresión 3D y se recomienda evaluar su escalabilidad y robustez con datos reales en entornos productivos.