Product design · Dylan Dotolo
Voice agent evals
Designing a way to turn voice agent behavior into checks that teams can score and review.
I explored an eval builder that keeps each criterion’s question, scoring method, and failure behavior together. The design uses a yes-or-no check for a greeting and session data for a turn-count check.
Creating an eval
The walkthrough names an eval, checks whether the agent greeted the caller, and adds a turn-count check. It shows how a criterion can be weighted to fail the session on its own.
Configuring a criterion
The same card adapts as the scoring method changes from yes or no to a scored scale or a session-data check.