My guess is that he is building the PDFs through latex/xetex. I built something similar to that, but instead of forking Pandoc (and investing the time to learn

My guess is that he is building the PDFs through latex/xetex. I built something similar to that, but instead of forking Pandoc (and investing the time to learn Haskell), I just added filters through Python. It's not at fast, but you can intercept mostly anything. Thus, the flow ends up as: paper.md -> Pandoc -> paper.json + metadata -> paper.tex -> paper.pdf