The complex topography of southern Chile causes high spatial variability in near-surface winds, challenging the resolution limits of current operational weather models. To address this, we propose a machine learning–based downscaling approach using gradient boosting (LightGBM) to generate high-resolution (333 m) wind forecasts from 3 km WRF outputs. The models incorporate meteorological, topographic, and wind–terrain interaction variables. Trained on 16 WRF cases and tested on 965 hourly maps, the approach achieved \(R^2\) values of approximately 0.88 (U10) and 0.93 (V10), outperforming previous methods. The framework offers a lightweight and operationally viable solution that can be transferred to similar applications. Enabling near real-time forecasts for maritime and aquaculture planning.