Getting Structured Data Out of Documents Nobody Agreed on a Format For | W. Balogun | SysConf 2026

Speaker: Wahab Balogun Wahab built a system that required extracting data from at least three sources—images, email, and PDF—and converged them into one structured extraction contract. These sources have different input shapes and distinct schemas, but must be inserted into the same ledger. The first part of addressing this involved OCR + LLM extractions. The second, and more critical, part was about ensuring the output is trustworthy enough to write to a system of record. This talk will dive into how Wahab implemented both parts for the pipelines he built.