Why Frontier Models Still Can’t Parse a PDF

Video summary

Reducto CEO Adit Abraham says its document AI processes pages for a penny, after surpassing 7 billion pages.

Reducto co-founder and CEO Adit Abraham tells a16z partner Jennifer Li that his company pivoted from AI memory to document processing after joining YC in 2023. Traditional OCR templates struggle with messy layouts, scans, charts, and spreadsheets, so Reducto combines computer vision and vision-language models and built its own R1 model. Its Deep Extract harness lets agents review and validate outputs, addressing errors that compound across multi-document workflows. Reducto charges a list price of one cent per page and had processed more than 7 billion pages; customers include hedge funds, private equity firms, and enterprises. Abraham says generic benchmarks can mislead, urging buyers to evaluate on representative data. He argues specialized document infrastructure can complement frontier labs, and describes the stakes: an OCR mistake can contribute to a rejected healthcare claim. He cites 80% of Medicare appeals overturning the original decision and says 16% of U.S. tax filers still submit paper returns.

Chapters

1. 0:00Pivoting at YC from memory to documents: Reducto’s 2023 shift to document infrastructure
2. 2:53From OCR templates to vision language models: W-2 templates give way to structured VLM parsing
3. 6:01Building R1, Reducto’s own model: production feedback tackles split tables and dense blueprints
4. 8:54Accuracy, compounding errors and Deep Extract: high-90s parsing and agent self-review
5. 11:19A penny per page and 7 billion pages processed: Reducto scales document workloads for finance
6. 15:14Why knowledge work lags behind coding: consequential decisions demand guardrails
7. 18:59Why the labs aren’t focused on this: document edge cases, recall failures, and invoice validation
8. 23:06Benchmarks have become marketing: Reducto tests noisy documents and trains chart agents
9. 29:58What Reducto becomes: reliable document tools for Harvey and 10 million-document workloads
10. 31:57When one OCR error denies a claim: sub-50% accuracy and overturned Medicare appeals