Nodal architectures have received an ever-increasing endorsement in computer graphics in recent years. However, creating a node-based system specifically tailored to GPU-centered applications with real-time performance is not straightforward. In this paper, we discuss the design choices we took in the making of Gratin, our open-source node-based system. This information is useful to graphics experts interested in crafting their own node-based system working on the GPU, either starting from scratch or taking inspiration from our source code. We first detail the architecture of Gratin at the graph level, with data structures permitting real- time updates even for large pipelines. We then present the design choices we made at the node level, which provide for three levels of programmability and, hence, a gentle learning curve. Finally, we show the benefits of our approach by presenting use cases in research prototyping and teaching.