They mention towards the end that they haven't found a natural programming model for this sort of thing yet and are playing around with different ideas. My initial thought was that an APL-style language with OCR could be a fairly natural fit, maybe as part of a node-based thing like their Crosscut project.