Shiv Indap of Upstart and Brooke Hopkins of Coval explain their voice-agent evaluation process at AI Engineer.
A narrow call with complicated rules
- The agent verifies identity and asks applicants about income continuity, intended use of funds and outstanding obligations.
- Indap says some responses require follow-up questions while others must not prompt them; compliance depends on context.
- The agent also sidesteps status and funding-timing questions outside its assigned task.
Human review and the AI judge
- Upstart initially reviewed every call by hand, but a shared spreadsheet became difficult to manage.
- An LLM judge now scores completed calls and explains its compliance assessment; reviewers check its output.
- Indap reports judge–human agreement rose from 33% to 92% after documenting unwritten rules, resolving reviewer disagreements and improving the judge.
- Those are the speakers’ internal agreement figures, not a measured compliance rate or proof that violations are caught.
Close the loop
- Define outcomes, required steps and conversation quality; revise metrics as the product changes.
- Route flagged calls to reviewers and confirmed problems to the appropriate engineering or compliance team.
- Re-simulate fixes with voice, accents, background noise and interruptions before exposing customers to them.
- Proposed automated fixes remain subject to human review; Upstart uses feature flags and compliance approval where needed.
“Evals are really the product definition.” — Brooke Hopkins