Recent advances in computational solvers combined with the introduction of exascale systems have greatly increased the resolution and fidelity of large-scale simulations. In parallel rapid progress and the adoption of deep learning have spurred the development of frameworks that integrate machine learning with scientific computing. We introduce a lightweight modular in situ coupling framework designed to leverage the Catalyst API to embed machine learning routines directly into simulation workflows. Our approach enables seamless interoperability between C++ and Python applications through the use of a solver-side data adaptor and a ParaView-based Python interpreter. We illustrate the framework's design and usability by instrumenting it within a proxy application of the HARVEY vascular flow solver. To demonstrate its practical utility we perform both training and inference of a point-cloud autoencoder entirely at runtime.
