Callum Ferris

Silent model updates break document pipelines without triggering alerts.

Missing table rows can pass validation while corrupting downstream analytics.

A standard protocol for document AI agents eliminates custom connectors.

Why accuracy, latency, and cost can't all be optimized together in document extraction.

Bad chunk boundaries tank RAG accuracy more than model choice.

Graph-based methods beat geometric sorting on complex document layouts.

Breaking documents into logical parts prevents extraction errors downstream.