← All speakers

Shiv Indap

Conference affiliation: Senior Engineering Manager · Upstart · 2026

Shiv Indap works on voice AI for lending operations and was listed as a senior engineering manager at Upstart for AI Engineer NYC 2026. His work addresses a specific problem in automated loan verification: an agent must understand an applicant’s answer well enough to determine which follow-up questions it is permitted to ask. A conversation can sound natural while taking a turn the lender’s rules prohibit.

On this page

From product engineering to lending automation

Shiv’s earlier work included lead development at Carsala, where his responsibilities spanned the company’s technology, and engineering leadership at Customer Lobby. His professional profile includes a recommendation describing the Carsala work. Customer Lobby’s team profile and a customer testimonial identify him as vice president of engineering at Customer Lobby, which builds customer-review and repeat-business software.

At Upstart, Shiv has worked with Coval to help the operations team scale AI workflows. In a public endorsement of the partnership, he credited Coval with helping that work grow. His joint work with Coval founder Brooke Hopkins connects voice-agent testing with human review of both the conversations and the automated systems used to assess them.

Testing the next question—and the evaluator

In their joint presentation on lending voice AI, Shiv and Brooke examined why conversation-level evaluation matters. An applicant’s explanation can change which questions are allowed next: a follow-up appropriate for graduation travel may be prohibited when the money is for medical expenses. Evaluating an isolated response misses that dependency. The test must account for what the caller has already volunteered and how the agent uses it to choose its next question.

Their collaborative workflow also subjects AI judges to human review. When a human reviewer and an automated evaluator disagree, the cause may be missing instructions, inconsistent human criteria, or an evaluator’s mistake. Investigating the disagreement determines what needs correction; treating every disagreement as an agent failure would obscure those differences. Confirmed failures feed further voice testing, with human approval required before changes reach applicants. The work extends evaluation beyond scoring a call to examining whether the scoring process itself applies the lender’s requirements correctly.

1 conference talk

Loading more talks and highlights…

References