Roadmovie was a research tool for project development, a space to load a corpus of documents, public funding databases, catalogues and PDF reports, and work with it through conversation. It was built in 2023 with developer Andrey Nikushkin, a year before Google launched NotebookLM and turned the same idea into a mass-market product.
The node canvas, with sources connected to a conversation and the documents being drafted. Images of the working process, reconstructed from the code. The application is discontinued.
The difference was the form. Where today’s tools put a sidebar, Roadmovie put a canvas in which sources were nodes connected to a conversation, and from that conversation came the documents being drafted. The research process was drawn as a graph you could read at a glance, which sources feed which questions and which texts have come out of them.
The test corpus was real, with the ICEC funding calls in PDF, the Spanish film catalogue database and tables from the Film Development Fund. Over those documents, a semantic retrieval system located the passages relevant to each question and the language model answered by citing them. The document view highlighted the passages used in each answer, and the drafting editor turned the conversation into a working text without leaving the tool.
The case that motivated it was a concrete one. Understanding how a film gets financed means reading through dozens of calls with different terms, deadlines that do not line up and criteria that change every year, and that complexity is paid for in hours of work. With the corpus loaded you could question the sources directly, what each call requires, which ones it can be combined with, what has been financed before and for how much, and put together viable financing models for a specific project.
Document and drafting, with the source and its quoted passages beside the draft.
Roadmovie ran on the model infrastructure of 2023, and that generation has been retired. When NotebookLM appeared with the same idea, backed by an infrastructure no production company could match, the decision was to close the development and keep what we had learned. Preserving it makes sense because it documents that the question was right, and that the node-canvas answer remains largely unexplored.
