Urban design necessitates accurate predictions of wind and contaminant dispersion to foster sustainable cities. Although computational fluid dynamics (CFD) models like large-eddy simulation (LES) offer precise predictions, they are computationally intensive. Steady Reynolds-averaged Navier-Stokes (SRANS) models are somewhat faster but suffer from inaccuracies due to the approximations involved. This study introduces a two-stage CFD-based artificial intelligence (AI) model that significantly reduces computational costs while maintaining high accuracy. The AI model employs a graph neural network (GNN) framework, using a SRANS model with a coarse grid as the initial state, and then refines the CFD outputs to achieve high-fidelity results in a single inference step. The GNN model was trained using data on airflow and contaminant dispersions across various scenarios with differing building structures and densities, incorporating corrections from key parameters derived from LES. When applied to predict airflow and contaminant dispersion in a section of Tokyo with available experimental data, the CFD-based GNN model produced results comparable to those of LES, while reducing computational costs by three orders of magnitude at the same grid resolution.