The paper reports work-on-progress on discovering undocumented links between source code updates and textual descriptions of desired features. The technical work is based on an existing trained model, T-BERT, that is shown to be effective for an example project by comparing the documented links to those automatically discovered. Prior work has concentrated on the development and evaluation of the ML model. Our aim is to take the next steps towards a practical tool. In this paper, we test the algorithm with a few example projects and analyze the discovered links.