× Dvar · brief
Prove every model on your own calls. Then upgrade.
Dvar makes the models you already run provable and replaceable: evals from your QA labels, checks around every model, a release gate on every change, and tuned open models on your GPUs. We start with one eval on outbound sales calls.
| # | Module | What Policybazaar gets | First result |
|---|---|---|---|
| 1 | Call summaries and QA | Every flag and disclosure traced to a line, eight checks per summary. | Week 3: an eval on one labelled slice of outbound sales calls |
| 2 | Advisor copilot and bots | Cited answers from the policy wording, checked before an advisor sees them. | Week 6: a regression suite on one bot |
| 3 | Claims and KYC documents | Claim files read field by field and checked against the policy. | Week 8: the harness on a sample of masked claim files |
| 4 | Evals and the release gate | Gold from your QA labels and a gate on every model change. | Week 2: first eval report on today's model |
| 5 | Model upgrades and inference | Newer models, tuned on your misses, at the lowest cost that passes. | Week 10: a tuned candidate through the gate |
What the demo shows
- • Call summaries and QA: 18 calls in three languages; four summarizers; flags linked to the transcript.
- • Advisor copilot and bots: 12 advisor questions; checks for promised returns, hidden conditions and uncited answers.
- • Claims and KYC documents: A five-document claim file, policy checks and a review queue.
- • Evals and the release gate: A scorecard on 18 calls; the Eval Console on 180.
- • Model upgrades and inference: The upgrade loop, cost at your volume, misses as training data.
To start
- • One workflow: outbound sales call QA.
- • 200 to 500 calls your QA team has reviewed, masked.
- • Today's model outputs on those calls, for the baseline.
- • An owner on your side and a place to run it.
Questions for the next call
- • How do you decide to swap a model today, and what would count as proof?
- • How many human-labelled calls a week exist for QA?
- • What does one analysed call cost you end to end today?
Concept demo by Dvar. Synthetic data. Timings are estimates.