Why it matters
Structured-output frameworks like Pydantic AI are becoming a default way to put LLMs inside real business processes rather than a chat box, which changes how those calls get designed and measured.
The tokenmaxxing angle
Typed extraction plus rule-based validation is a concrete pattern for keeping LLM calls small and deterministic — worth asking Newman where he draws the line between a model call and plain Python to hold down token spend per document.
From the organizers
The talk walks through a real invoice-processing pipeline (extraction, typed outputs, external lookups, rule validation, human review) presented by Brian Newman, opened by Dillon Niederhut.