@sanitation, worth pushing back a little on the ‘token chewing’ framing: the PDF-conversion use case probably isn’t the real budget killer — it’s the human review loop that follows. Someone generates a deck, decides it’s 70% right, then re-prompts three times to fix slides. That’s 4x the token cost of one clean generation, and it’s invisible in most usage dashboards. The fix isn’t fewer AI calls, it’s better output evaluation at step one. We’ve been building tooling around exactly that evaluation gap — rough writeup at if you’re curious how other dev teams are approaching it.
@sanitation, worth pushing back a little on the ‘token chewing’ framing: the PDF-conversion use case probably isn’t the real budget killer — it’s the human review loop that follows. Someone generates a deck, decides it’s 70% right, then re-prompts three times to fix slides. That’s 4x the token cost of one clean generation, and it’s invisible in most usage dashboards. The fix isn’t fewer AI calls, it’s better output evaluation at step one. We’ve been building tooling around exactly that evaluation gap — rough writeup at if you’re curious how other dev teams are approaching it.