KreditBee × Dvar

How Dvar helps KreditBee

Dvar does three things: evals that gate every AI change, inference served from India, and post-training of models KreditBee owns. Here is where each fits, against what KreditBee has said in public.

What KreditBee needs (public sources)	Where Dvar fits	See it
A calling agent for reminders and collections, now with GreyLabs (as we understand it), earlier with Gnani. Voice AI for pre-due and post-due reminders across millions of borrowers, integrated with the LMS and CRM; the vendor reports 46% lower opex and 75% connectivity. Gnani.ai case study, undated GreyLabs AI builds multilingual voice agents for sales, service, collections and renewals in BFSI; ₹85 Cr Series A led by Elevation Capital. Z47, 6 Oct 2025	A vendor-neutral release gate: every agent version, prompt or model change played against KreditBee's scripted borrowers before it goes live, and every live call scored, not a 5% QA sample.	Voice agent gate →
Collections at scale: a 1,800-person in-house team and a high on-time rate to protect. In-house collections team of about 1,800; about 93.5% of repayments on time, nearly 99% within 30 days. Inc42, Apr 2026	A post-call copilot on a small model: a disposition for every call, with stop requests, hardship and third parties caught first.	Collections copilot →
Underwriting on data from about 150 sources, with bank statements through Account Aggregator. About 70 lakh customer interactions handled with AI-assisted systems in FY26; data from about 150 sources; 8 Cr applications underwritten. Inc42, Apr 2026 Bank statements for underwriting pulled through the Account Aggregator framework with Digitap; over 1 lakh in the trial. CXOToday, 10 Mar 2023	Statement lines labelled by a small model and the memo computed in code, so a label error is visible and scored. The model never takes the credit decision.	Underwriting copilot →
KYC and income documents for 30,000 loans a day. About 30,000 loans a day; 18 million unique borrowers; 60 million loans disbursed; about 70,000 app downloads a day. Inc42, The Making Of A Lending Tech Unicorn, Apr 2026	One harness for every document: a small vision model first, a large one only for the hard field, checks a model cannot skip, a review queue that becomes training data.	KYC documents →
“Embedding AI into the core of our lending stack”, and hiring for LLM features and model evaluation. CEO Madhusudan Ekambaram: “embedding AI into the core of our lending stack” for underwriting, risk controls and a faster customer experience. Business Today, 8 Apr 2026 A Generative AI intern role: build, prompt and evaluate LLM features with Python, LangChain, RAG, vector databases and model evaluation. Job listing, 2026	Small models post-trained on KreditBee's own traces and reviewer corrections. KreditBee owns the weights and the names; each release goes through the same gate.	Custom models →
Cost discipline on the way to an IPO. Standalone FY25: profit up about 87% to ₹237 Cr while total expenses fell over 31%. MediaNama, Jan 2026 Raised $280M in a Series E at a $1.5B valuation; preparing for an IPO; plans to scale AI for credit risk and customer targeting. MediaNama, 8 Apr 2026	Inference served from India at small-model prices, each task routed to the cheapest model that passes its checks, with a frontier model as the fallback.	Inference cost →

How the three work together

Evals find the misses
Misses become training data
KreditBee custom models
Served on Dvar inference
Every call traced and scored again

Beside the voice vendor, not instead of it

• Dvar reads the agent's transcripts and tool events from the vendor's export or webhook. Nothing changes in the call path.
• Before a release, the vendor's staging agent takes the scripted borrowers; KreditBee sees the gate result, not the vendor's word.
• After a release, every live call is scored on the same checks, so a drift shows up in days.
• The checks are KreditBee's: contact hours, identity, privacy, no coercion, hardship, stop requests, languages.

A four-week pilot

Week 1. Pick two steps (the calling-agent gate and post-call dispositions are our suggestion). Agree the checks and critical failures with collections and compliance.
Week 2. Build the test suite and gold sets from redacted calls and synthetic borrowers in Hindi, Kannada, Tamil, Telugu and English.
Week 3. Run the current agent, a Dvar-hosted small model and a frontier model through the same scorer.
Week 4. Read the results together and decide per step: keep, switch, or route the hard cases. The suite stays as the regression test.

Sources read on 22 Oct 2026. The GreyLabs calling agent is as we understand it from KreditBee; we did not find a public source for it. Everything in this demo runs on synthetic data; no KreditBee data was used.