i actually expect that this will work pretty well, at least for academic articles, because the machine learning engineers working on it will constantly be "dogfooding" it, as machine learning research is published mostly as PDFs.