In April 2025, Thomas Blume developed a major feature for the Nixtla/neuralforecast repository, focusing on enhancing the NBEATS model with dynamic basis functions. He architected updates that allow the model to integrate multiple basis types, such as Legendre, polynomial, and changepoint, improving adaptability to diverse time series patterns. Using Python, PyTorch, and NumPy, Thomas designed and validated the integration to support more flexible forecasting pipelines. This work addressed the challenge of static basis configurations, enabling more accurate forecasts with less manual tuning and supporting business needs like inventory optimization and demand planning through improved model flexibility and accuracy.