Digital Twins are transforming engineering and industrial sectors by enabling real-time monitoring and predictive control of complex dynamical systems. However, their reliability hinges on the capability to maintain synchronization between the virtual model and the physical entity, a significant challenge in chaotic scenarios such as fluid sloshing. While traditional computational methods struggle with the efficiency required for real-time feedback, and pure data-driven models often lack physical interpretability, a hybrid approach is required to handle the high dimensionality and non-linear nature of free-surface flows. In this work, we present an enhanced Digital Twin framework for fluid sloshing that couples Thermodynamics-informed Graph Neural Networks (TIGNNs) [1] with a computer vision-based data assimilation system. Our approach leverages the GENERIC formalism [3] to ensure the physical consistency of the fluid dynamics, while simultaneously integrating a semantic segmentation pipeline to process real-time video feeds. This setup establishes a bidirectional information flow: the computer vision system extracts the instantaneous topology of the real fluid to correct the neural network's inference and prevent drift, while the physics-based model resolves and visualizes non-observable state variables, such as internal energy and velocity fields, onto the digital replica. The integration of the computer vision algorithms allows the Digital Twin to "see" the physical object, using geometric data to constrain and refine the thermodynamic simulation in real-time. This methodology not only significantly reduces computational costs compared to traditional CFD solvers but also ensures that the Digital Twin is not merely a simulation, but a faithful, self-correcting mirror of reality. The result is a robust tool capable of predicting and plotting complex fluid behaviours under dynamic excitations, bridging the gap between visual observation and physical inference.