Evaluating and Controlling AI Hallucinations

Ex-Google Duplex Eng Lead on Disrupting $2B Clinical Trials with AI - Patrick Leung

Topic 5 of 18

Published Jul 28, 2025 · Clip 10:11–14:30 (4:18)

Patrick describes how Faro moved from naive prompt tweaking to a systematic evaluation architecture. Each generated subsection is checked against specific checklists for factual content, tone, style, and detail, evolving into an agentic architecture with multiple evaluation agents feeding feedback into regeneration cycles. He also explains the human-in-the-loop bootstrapping process: engineers, data scientists, clinical writers, and QA initially examine everything, then codify checks into automated evaluation until reaching confidence across therapeutic areas before sharing with early adopter customers.