Policybazaar × Dvar
How Dvar helps Policybazaar
The eval and release layer around the AI you already run: your evals, your models, your GPUs.
Evals
Know a model is right before it ships.
- • Gold from your QA reviews
- • Code checks, critical ones marked
- • A judge calibrated to your reviewers
- • A gate on every change
Harnesses
The same checks around every model.
- • Checks before an answer reaches anyone
- • Small model first, large only on a miss
- • Review queues that feed training data
- • Identifiers masked at entry
Tuned models and inference
The lowest cost that passes the gate.
- • Tuned on your reviewed misses
- • You own the weights
- • On-prem or your cloud
- • Cost per call, not per token
Where it fits what you run today
| Workload | Today | With Dvar |
|---|---|---|
| Call summaries and QA | ~21 lakh calls a day, 100% read for quality, mis-selling and data leakage | Every flag and disclosure traced to a line, eight checks per summary. |
| Advisor copilot and bots | ~1 lakh advisor assistant queries a day; 35 voice and chat bots live | Cited answers from the policy wording, checked before an advisor sees them. |
| Claims and KYC documents | ~7 lakh documents a day; agentic claims checks; issuance in 15 minutes for half of customers | Claim files read field by field and checked against the policy. |
| Evals and the release gate | The stated blocker: no evals to trust a model swap | Gold from your QA labels and a gate on every model change. |
| Model upgrades and inference | Open models fine-tuned in-house; GPUs mostly on-prem; outcomes over token counts | Newer models, tuned on your misses, at the lowest cost that passes. |
How it works together
- Evals find misses
- Reviewers confirm
- Misses train the model
- The gate proves it
- It ships
Start: one workflow, one labelled slice, first eval report in two weeks.