Our platform employs geometric deep learning and molecular modelling for rapid, high-throughput virtual screening, analyzing billions of molecules within hours from a database subset representing a 0.001% of the total library. It uses this data to train a surrogate model that estimates docking scores across the entire database, validated through an induced fit docking Monte Carlo algorithm. With an active learning strategy designed for efficient retraining, the system thoroughly navigates extensive molecular spaces. An explainability component underscores key docking substructures, facilitating lead optimization, which is later performed by a pocket geometry constrained diffusion model.