Daily rainfall forecasting in Assam, northeast India, remains challenging due to strong monsoon variability and complex terrain. This study compares Multiple Linear Regression (baseline), XGBoost, and LightGBM for daily precipitation prediction across five hydro-climatic regions of Assam using a 24-year meteorological dataset (2001–2024) from IMD (India Meteorological Department) and NASA POWER. Models were trained on 2001–2023 and evaluated on an independent 2024 test year. Both gradient boosting models substantially improved prediction accuracy relative to linear regression, achieving R² (Coefficient of Determination) values of 0.775–0.974 and reducing mean absolute error by up to 85%. Event-detection skill was also strong, with CSI (Critical Success Index) values of 0.852–0.958 at the 5 mm/day rainfall threshold and consistent detection of heavy rainfall events at 25 mm/day, where linear regression showed very limited skill. Seasonal analysis indicated higher uncertainty during the peak monsoon season, particularly in the orographically complex Upper Assam region where LightGBM monsoon R² declined to 0.687. The analysis is limited by the use of single representative grid points per region and evaluation on a single test year, which may not fully capture inter-annual variability. Nevertheless, the results suggest that gradient boosting models with temporal feature engineering and strict train–test separation provide a promising framework for operational rainfall forecasting in monsoon-dominated regions such as Assam.