Dvar for KreditBee: one page

The idea

KreditBee already runs AI from KYC to collections, at about 30,000 loans a day. The risk at that volume is not one bad answer; it is a change that quietly breaks a rule on thousands of calls. Dvar puts a release gate on every AI step, runs the routine traffic on small models served from India, and post-trains models KreditBee owns on its own misses.

What the demo shows

• Voice agent gate: a candidate agent that fixes the live version's problems but tells a spouse the amount due is blocked; the fixed version ships.
• Collections copilot: dispositions for every call, with hardship, stop requests and third parties caught first, scored against an answer key.
• Underwriting copilot: statement lines labelled by a small model, the memo and decision computed in code.
• KYC documents: a small model first, checks it cannot skip, a review queue that becomes training data.
• Custom models and cost: misses become training examples; monthly cost at lending volume.

A four-week pilot

Week 1, agree two steps and their checks. Week 2, build the suite and gold sets from redacted calls and synthetic borrowers. Week 3, run the current setup, a Dvar-hosted small model and a frontier model through one scorer. Week 4, decide per step: keep, switch, or route the hard cases.

What we need from KreditBee

• An owner in collections or the AI team to agree the checks and the critical failures, with compliance.
• Read access to calling-agent transcripts and tool events (redacted), or a staging endpoint for the vendor's agent.
• A redacted sample of call dispositions and statement labels, and the current model's outputs on them if you can share them.
• Where to run Dvar Evals: your cloud, or ours in India with synthetic data first.