Why 99% PDF Parsing Accuracy Still Isn’t Enough in Production

Reducto processes a billion PDF pages a month. Co-founder and CEO Adit Abraham went on the Latent.Space podcast and, while making steak sandwiches with Allen Park, talked about why PDFs are still tripping up AI to this day.

On the show, Adit described a problem that is devastating in production but almost invisible in demos: frontier models silently drop rows. A table is missing a row, headers are merged incorrectly, and the model doesn’t report an error; its tone remains just as confident. You get a result that looks complete—until finance finds that the missing row is exactly the one that mattered.

A commenter added a sharp point: the 99% figure itself is a trap. He only trusts results after adding a row-count check and checksum to table and invoice extraction. Another person mentioned a study of 8,000 academic PDFs from four major academic publishers: only 2 had tagged tables. In other words, the file itself almost never tells the model 'this is a table.' The model has to guess.

Adit’s solution is not to bet on a single model. Reducto uses model routing: difficult pages and strange tables go to models better suited to them, while ordinary pages are kept cost-controlled. The show also breaks down when to train your own model versus when to buy one. He discussed the accuracy-versus-cost trade-offs of Reducto R1, and how at a scale of a billion pages a month, 99% accuracy simply isn’t enough.

“Everyone can win their own benchmark” was another line from Adit. That’s why Reducto relies more internally on continuous evals and production failure cases than on leaderboards. They also use AI agents for internal work, and the hiring process keeps an eight-hour onsite to see what early employees’ kind of “obsession” looks like.

The podcast ends by discussing office work’s Claude Code moment: from reading PDFs to redlining contracts to extraction and citation. Adit thinks there’s still a way to go, but the direction is clear.

Midway through the show, Adit torches peaches, makes chimichurri, and assembles steak sandwiches. A commenter recognized it as a South African prego roll and said it makes perfect sense to eat this at 1 a.m., with burrata counting as a fair upgrade.

There’s also a segment in the video about a steak “accusation” and a taste test, placed at the end.

Original link: https://x.com/latentspacepod/status/2118800336412709024