Voice AI Where One Wrong Question Breaks the Rules - Shiv Indap, Upstart & Brooke Hopkins, Coval

Upstart, a lender, uses voice AI to automate loan verification calls, and it handles thousands of lending calls a day. In a talk published by the AI Engineer channel, Shiv Indap, who leads voice AI at Upstart, and Brooke Hopkins, founder of Coval and former head of evaluation infrastructure at Waymo, explain how they check those calls when, as the description puts it, one wrong question can become a regulatory problem.

The description says human review of calls first lived in a Google Sheet, which the team outgrew. Checking at volume relies on an LLM judge: a language model that reads a call and scores it against defined metrics, standing in for a human reviewer. A judge is only useful if its verdicts match what human reviewers would decide, and the talk reports raising agreement between the judge and human reviewers from 33% to 92%.

The judge sits inside a larger loop that, per the description, runs from monitoring live calls and queuing them for human review to re-simulating a fix before it reaches production, and on to self-healing agents. Every change is still signed off by a human. This account comes from the publisher's description and chapter list; no transcript was available, so it cannot say how the agreement was raised, how it was measured, or what the self-healing agents change on their own.