Policybazaar × Dvar · concept demo

Prove every model on your own calls. Then upgrade.

You already read every call. Dvar adds the evals that let you trust what the models say, swap in newer ones when they win, and see the cost of every analysed call: your evals, your models, your GPUs.

Start the guided tour Open the call QA desk How Dvar helps Policybazaar

checks on every call summary

Figures from the call, no mis-selling missed, no identifier repeated, and five more

critical misses for the open model + examples

On the 18 sample calls, measured on real model calls; the open model alone had 2

languages, one report

English, Hindi and Hinglish calls, summarized in English for QA

Five modules · about 15 minutes to see them all Step-by-step checklist

Module 1Call summaries and QAToday: ~21 lakh calls a day, 100% read for quality, mis-selling and data leakageSummaries that QA and the next agent can trust: every flag and disclosure traced to a line, eight checks on every summary.Open the demo →Module 2Advisor copilot and botsToday: ~1 lakh advisor assistant queries a day; 35 voice and chat bots liveAnswers from the policy wording, cited, with runtime checks and a closed-source fallback; the same harness and regression suite for every bot.Open the demo →Module 3Claims and KYC documentsToday: ~7 lakh documents a day; agentic claims checks; issuance in 15 minutes for half of customersClaim files read field by field, checked against the policy, routed small to large to a person, with every decision traced.Open the demo →Module 4Evals and the release gateToday: The stated blocker: no evals to trust a model swapThe Dvar Eval Console on your traffic: gold from your QA labels, a calibrated judge, a gate on every model or prompt change.Open the demo →Module 5Model upgrades and inferenceToday: Open models fine-tuned in-house; GPUs mostly on-prem; outcomes over token countsA repeatable path to newer models, tuned on your misses, at the lowest cost per call that passes the gate, on your GPUs.Open the demo →One-page brief →Prints to PDF

Words used in this demo

Today's model: A stand-in for the summarizer in production today. In a pilot, we measure the real one on your calls before anything else.
Closed-source model: A model whose inner workings are kept private by the company that makes it, such as Google, Anthropic or OpenAI. You can only use it through their paid service.
Open model: A model whose weights are published, so it can run on Dvar in India or on your own GPUs.
Answer key (gold): The reviewed summary of a call. In a pilot it comes from the reviews your QA team already does.
Critical check: A check that blocks a model on a single miss: an invented figure, a missed mis-selling flag, a repeated identifier.
Release gate: The rule a model must pass before it takes traffic: no critical miss, no drop on any topic, within the time budget.
Tuned model: An open model trained further on your reviewed calls. You own the weights; it runs on your GPUs. Not in this demo yet: its “open model + examples” puts reviewed summaries in the prompt instead.
Judge: A model that scores answers automatically, trusted only after it agrees with your reviewers on a labelled set.

Everything here is synthetic and built by Dvar; no Policybazaar data was used. Volumes quoted are from PB Fintech's public Q1 FY27 investor deck. The insurers and products in the calls (Sampada Life, Arogya Plus, Suraksha General) are made up.