FloodCast is a novel framework for large-scale flood modeling and forecasting, combining multi-satellite observations with a geometry-adaptive physics-informed neural solver (GeoPINS). The framework includes a multi-satellite observation module and a hydrodynamic modeling module. The observation module uses real-time unsupervised change detection and a rainfall processing tool to leverage multi-satellite data for flood prediction. The hydrodynamic module employs GeoPINS, a physics-informed neural solver that adapts to complex river geometries and is resolution-invariant. GeoPINS is based on Fourier neural operators and is trained without requiring training data, enabling accurate and efficient flood simulations. A sequence-to-sequence GeoPINS model is developed to handle long-term temporal series and large spatial domains. The framework was tested on the 2022 Pakistan flood, using SAR-based flood data, traditional hydrodynamic benchmarks, and optical remote sensing images. Results showed that GeoPINS outperformed traditional hydrodynamic models in predicting flood inundation depth and extent, with an average MAPE of 14.93% and MAE of 0.0610m for 14-day water depth predictions. FloodCast is publicly available at https://github.com/HydroPML/FloodCast. The framework enables high-precision, large-scale flood modeling and real-time flood hazard forecasting using reliable precipitation data.FloodCast is a novel framework for large-scale flood modeling and forecasting, combining multi-satellite observations with a geometry-adaptive physics-informed neural solver (GeoPINS). The framework includes a multi-satellite observation module and a hydrodynamic modeling module. The observation module uses real-time unsupervised change detection and a rainfall processing tool to leverage multi-satellite data for flood prediction. The hydrodynamic module employs GeoPINS, a physics-informed neural solver that adapts to complex river geometries and is resolution-invariant. GeoPINS is based on Fourier neural operators and is trained without requiring training data, enabling accurate and efficient flood simulations. A sequence-to-sequence GeoPINS model is developed to handle long-term temporal series and large spatial domains. The framework was tested on the 2022 Pakistan flood, using SAR-based flood data, traditional hydrodynamic benchmarks, and optical remote sensing images. Results showed that GeoPINS outperformed traditional hydrodynamic models in predicting flood inundation depth and extent, with an average MAPE of 14.93% and MAE of 0.0610m for 14-day water depth predictions. FloodCast is publicly available at https://github.com/HydroPML/FloodCast. The framework enables high-precision, large-scale flood modeling and real-time flood hazard forecasting using reliable precipitation data.