Many regions, especially in countries like India, lack reliable, high-resolution data on daily baseflow because traditional methods require continous streamflow data, and dense river monitoring networks. In this study, we address this challenge by creating the first high-resolution daily baseflow dataset for hundreds of river catchments across India, covering the period from 1980 to 2020. We used data from the CAMELS-IND database and applied a cutting-edge deep learning model known as the Temporal Fusion Transformer (TFT). This model can learn complex patterns by combining catchment features like topography and land use with daily weather data such as rainfall and temperature. We grouped similar catchments using hydrologic clustering techniques and used twelve traditional methods to help train and evaluate our model. Our results show that the TFT model provides accurate and consistent baseflow estimates across India’s diverse climate zones, even in areas with limited data. This new dataset can help researchers better understand water availability, track long-term changes in groundwater, and support planning for droughts and floods. It also demonstrates the power of advanced machine learning tools to solve critical water challenges in data-limited regions.