Layered Fact Validation with LLM-as-Judge

LLMs hallucinate even when you ground them in the facts. In regulated domains, a single invented detail can invalidate an entire output. This talk explores how to catch hallucinations before they reach users, covering claim decomposition, per-claim verification, and explainable failure reporting.

We'll look at why a layered approach matters: combining cheap deterministic checks with an LLM judge and the tradeoffs of each. The talk also covers judge bias and calibration, confidence-gated decision policies, and graceful degradation so verification never takes down generation. Small pattern, big payoff: trust.

👨🏻‍💻 SPEAKER: Percy Brea

🗓 AGENDA:
6:30 PM: Arrival /Networking
6:40 PM: Announcements by organizers
6:50 PM: Layered Fact Validation with LLM-as-Judge
8:15 PM: End of meeting
8:30 PM: Exit to after party celebration at Patio de Leon (courtyard behind Downtown House of Pizza)

👾 JOIN US ON DISCORD

SWFL Coders, a founding member of the Tech Alliance of SWFL, shares a Discord server with other local tech groups.
Click here to join the server and look for #💻┃swfl-coders, the channel dedicated to our group's communication.