Concept demo for KreditBee, by Dvar

Every AI step in the lending journey, measured before it reaches a borrower

KreditBee runs AI from the first KYC document to the last collections call, at about 30,000 loans a day. Dvar gives each step a release gate, runs the routine traffic on small models served from India, and post-trains models KreditBee owns.

Start the tour (9 stops)How Dvar helps KreditBee

Synthetic data throughout: 15 scripted calling-agent scenarios, 160 collections calls, 28 bank statements and five specimen KYC documents. Model results are an internal benchmark on synthetic data.

Voice agent release gate

Before a new version of the AI calling agent goes live, it is tested on practice calls in five languages. If it breaks even one hard rule, such as telling a relative how much is owed, it is held back.

Collections copilot

After each collections call, a model notes what the borrower said: a promise to pay, why they are late, money trouble, a request to stop calling. Simple rules then pick the next step.

Underwriting copilot

A model sorts each line of a bank statement: salary, other loans, bounced payments. Income and the loan decision are then worked out by plain code, not by the model.

KYC documents

A model reads ID and income documents. Automatic checks catch its mistakes, and anything doubtful goes to a person.

Custom models

The model's mistakes become examples to train a better model, one that KreditBee owns.

Inference cost

What the AI would cost each month at KreditBee's volume, worked out from the measured run.

What this is not: KreditBee's product, data, models or calling agent, or a claim about how KreditBee builds these today. It is a generic lending workflow on synthetic data, built to show how each step can be run and measured.