Don't Approve the AI Model. Approve the App - Doneyli De Jesus, Langfuse
AI Engineer — Nov 4, 2026 — talk
Why watch
Doneyli De Jesus from Langfuse explains how banks implement deployment gates for AI agents by certifying applications rather than models. The talk covers building golden datasets, scoring traces for code quality and correctness, and using LLM judges with domain expert thresholds to evaluate AI outputs. A key example shows a small model with 97% grounding but 60% numeric accuracy passing an 85% accuracy gate after receiving a calculator tool. This approach reduced evaluation time from weeks to minutes.