This work introduces CEopt [1], a MATLAB-based tool leveraging the Cross-Entropy (CE) method, tailored for model calibration, updating, and data-driven modeling within scientific computing and reduced-order modeling. The CE method [2], renowned for its efficiency in handling complex, nonconvex optimization problems, forms the core of CEopt. It converts deterministic problems into stochastic estimation tasks, simplifying them through adaptive importance sampling that iteratively refines probability distributions to focus on promising regions. This approach makes CEopt particularly adept at calibrating complex models within machine learning frameworks, where traditional optimization tools may struggle. CEopt's "gray-box" solver offers transparency and intuitive control over optimization parameters, enhancing user understanding and interaction with the tool. It robustly handles both equality and inequality constraints, essential for the detailed demands of scientific workflows and the precise training of models under conditions of uncertainty. Through practical case studies, we illustrate how CEopt supports the integration of advanced machine learning techniques into scientific workflows, facilitating the development of efficient surrogate models and digital prototypes. The tool's flexibility and power to adapt to various scientific computing needs highlight its significance, particularly in enabling technologies crucial for scientific machine learning and reduced-order modeling. This work details CEopt's capabilities and potential in bridging the gap between high-level machine learning applications and practical, effective optimization solutions.