Voice AI Where One Wrong Question Breaks the Rules - Shiv Indap, Upstart & Brooke Hopkins, Coval
Key Takeaway
Integrating human review queues with automated LLM judges and voice-to-voice simulations increases compliance agreement from 33% to 92% for regulated voice agents.
Highlights
Upstart automates 10% of loan verification calls that are not instantly decisioned by using voice agents to ask three specific applicant questions.
Upstart's compliance agreement metric between human reviewers and LLM judges increased from 33% to 92% after encoding tribal knowledge and refining the evaluation process.
Autonomous vehicle evaluation infrastructure principles from Waymo apply directly to building reliable multi-model autonomous voice agent systems.
Voice agents in lending must handle strict regulatory constraints, such as prohibiting loan amount inquiries for medical fund usages.
Coval provides automated voice-to-voice simulation evaluations to test background noises, accents, and interruptions before deploying fixes to production.
Timeline
Voice Agent Use Case in Regulated Lending
Founders from Koval and Upstart discuss how evaluation infrastructure transfers from autonomous driving at Waymo to enterprise voice agents. Upstart deploys voice agents to handle verification interviews that require asking about income continuity, fund usage, and outstanding obligations. Because lending is heavily regulated, conversational mistakes carry severe legal consequences alongside customer experience issues, necessitating precise behavioral boundaries.
- Autonomous voice agents share architectural similarities with autonomous vehicles by requiring reliable multi-model reactions across complex paths.
- Upstart automates verification calls for the 10% of loan applicants who do not receive instant decisions.
- Regulatory constraints dictate specific rules, such as omitting loan amount questions when applicants specify medical fund usage.
Human Review Infrastructure and LLM Judges
Initial manual call auditing relied on basic spreadsheets that quickly fragmented under expanding team inputs. To scale oversight, automated LLM judges evaluate calls upon completion, generating a yes-or-no compliance score alongside an explanatory rationale. Continuous comparison against expert human reviews uncovered gaps driven by unwritten tribal knowledge, which engineers subsequently encoded into the system to dramatically improve judge reliability.
- Initial human review processes using Google Sheets broke down as operational teams scaled custom columns.
- Automated LLM judges evaluate completed calls against criteria like accuracy, clarity, completeness, and compliance alignment.
- Alignment between human expert reviews and LLM judge outputs rose from 33% to 92% through iterative codification of tribal knowledge.
Continuous Improvement Loop and Voice Simulations
Effective agent improvement requires a closed-loop operational practice that routes flagged failures to appropriate engineering or compliance teams. Instead of risking user experience by testing fixes directly in production, teams utilize voice-to-voice simulation platforms like Coval. These simulations test environmental variables such as accents and interruptions to verify bug remediation before feature flags release updates to live borrowers.
- Evaluation infrastructure functions as product definition by clarifying exact operational parameters and boundaries.
- Live call monitoring detects known failure patterns and identifies unknown anomalies through clustering analysis.
- Voice-to-voice simulations test prompt fixes against background noise and accents before updates impact production users.
Description
Upstart handles thousands of lending calls a day, where one wrong question can become a regulatory problem. Shiv Indap, who leads voice AI at Upstart, and Brooke Hopkins, founder of Coval and former head of evaluation infrastructure at Waymo, walk through automating loan verification calls: the compliance traps, outgrowing a Google Sheet for human review, and raising agreement between the LLM judge and human reviewers from 33% to 92%. Then they lay out the full loop from live monitoring and human review queues to re-simulating fixes and self-healing agents, with every change still signed off by a human. Chapters: 0:00 Intro 1:02 Voice AI at Upstart 2:11 Automating verification calls 3:04 Where compliance gets tricky 4:14 Outgrowing the Google Sheet 5:28 Metrics and LLM judges 7:03 From 33% to 92% judge agreement 8:52 The continuous improvement loop 10:04 Defining metrics 10:51 Monitoring live calls 12:31 Re-simulating fixes before production 14:04 Self-healing agents 15:11 Wrap-up: humans stay in the loop Shiv Indap https://www.upstart.com Brooke Hopkins https://www.coval.dev Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!