“Strong team, but small-data PDF extraction was hard”

I worked on Google Cloud Document AI for about 1.5 years. I ran experiments on extracting structured data from PDFs, including whether the model could train from very small document sets. Those tests produced a durable conclusion for me: it usually needed around 100 documents. My broader learning was that even a large engineering effort did not easily push the system past about 98 F1. The team and my manager were good, but for messy bespoke PDFs the product felt less magical than buyers might expect.

🔗 https://cloud.google.com/document-ai

• No comments yet.