This study aims to improve local-scale climate simulations by developing a Multi-Model Ensemble using selected General Circulation Models (GCMs). Bayesian Networks were applied to determine the optimum number of GCMs for simulating daily and monthly precipitation, as well as maximum and minimum temperatures (Tmax and Tmin), across 35 major cities in India, representing various meteorological sub-divisions. Five state-of-the-art machine learning models, including Random Forest (RF), Multivariate Adaptive Regression Splines, Support Vector Regression, Extreme Gradient Boosting (XGBoost) and Stacking, were used for these simulations. The study proposes a novel two-step approach for modeling daily precipitation which improved correlation coefficients (R) from 0.1 to 0.4 in low rainfall regions and from 0.4 to 0.52 in high rainfall regions. For monthly precipitation, RF and Stacking achieved R values ranging from 0.05 in arid regions to 0.93 in high-rainfall areas, with NRMSE values as low as 37.1%. Temperature simulations performed consistently well, with R values up to 0.99 for monthly Tmin, and daily Tmax simulations showing Normalized Root Mean Square (NRMSE) improvements from 16.8 to 60.4%. The Modified Degree of Agreement (MD) for low rainfall regions improved from 0.3 to 0.67 using RF. This study addresses critical gaps in GCM selection and rainfall simulation, demonstrating that the combination of Bayesian Networks and the two-step precipitation model enhances the accuracy of climate predictions at daily and monthly scales, especially across regions with varying rainfall patterns. Given the importance of precipitation as a variable, and based on the superior performance of RF and Stacking in simulating daily precipitation across the 35 locations, these models were selected for the nationwide analysis.