Climate change is disproportionately affecting vulnerable communities in the Global South, highlighting the need for updated building regulations to address overheating risks. To this end, this study presents a procedure for developing guidelines for climate-resilient residential building design. It compares future climate projections for five Latin America cities and generates urban weather files by integrating IPCC emission scenarios with urban heat island (UHI) effects. The research compares Extreme Gradient Boosting (XGBoost) and Artificial Neural Networks (ANN) as surrogate models, utilizing transfer learning techniques to efficiently evaluate three architectural designs. Following this, these surrogate models are combined with the Non-Dominated Sorting Genetic Algorithm III (NSGA-III) to improve thermal comfort and decrease energy demand. The procedure incorporates an optimization loop, cluster analysis, heatmaps, and frequency analysis to identify design thresholds for transparent and opaque building components. The results revealed a strong similarity in climate projections when comparing Special Report on Emissions Scenarios (SRES) and Shared Socioeconomic Pathways (SSP) scenarios, with future urban weather conditions trending towards higher dry-bulb temperatures, relative humidity, and global horizontal radiation. The performance of XGBoost and ANN was comparable, with transfer learning proving to be beneficial. Design thresholds were established for three key building components: walls, roofs, and windows; with certain configurations consistently emerged as favorable, including low solar absorptance, small and shaded windows, and low thermal transmittance and thermal capacity for roofs. Finally, these findings underscored the potential of this procedure to provide valuable insights for policymakers in Latin America.