Recent advancements in machine learning and observation availability have enabled data-driven flow field reconstructions. This study applies SHRED (Shallow Recurrent Decoders) to high-resolution coastal ocean dynamics, demonstrating the fidelity possible using existing time-series from wave buoys, CTDs, and water level sensors. The Salish Sea, an important urban sea system, serves as the testbed. Existing operational models like SSCOFS, LiveOcean, and SalishSeaCast allow for rigorous comparison, showing that while sea surface elevation is easily reconstructed, variables like salinity prove more complex. This region benefits from rich observations, including the VENUS observatory, which provides necessary current and salinity measurements. SHRED represents complex flow via nonlinear modes predicted from time-series measurements. Training utilizes 3D data from SalishSeaCast, with validation using observed time-series of essential ocean variables. These applications enable straightforward prediction of physical time-series. Moving from point validation to full field reconstruction via SHRED, this method integrates observed data with modeled physics. This offers more than simple interpolation, providing a foundation for future digital twins of the coastal ocean.