Your real traffic becomes the test
We build the eval with your team, so quality is quantified.
India's gateway to intelligence
Models built around your work.
Enterprise-specific intelligence, trained on your evals. We post-train until it beats a frontier API on your work, at up to 99% lower cost, with data staying in India.
Frontier research built around your business
One measured workflow carries from customer traffic to a model your team can own.
WORKFLOWReal traffic
EVALUATIONYour quality bar
MODEL SEARCHBest-fit model
POST-TRAININGAdapt to the eval
DEPLOYMENTRun in India
Find the model that fits
We compare open and frontier models on the same eval, then choose the lowest-cost one that clears your quality bar.
We post-train the winner
We post-train the best open model to give you the best cost per outcome. Custom-built for your workflow, not a shared API.
Production closes the loop
It runs on sovereign India servers on Dvar's platform, with production feedback continuously improving it.
The build loopOne workflow · four steps
- 01EVALUATE
- 02SELECT
- 03ADAPT
- 04DEPLOY
01Your real traffic becomes the test
We build the eval with your team, so quality is quantified.
02Find the model that fits
We compare open and frontier models on the same eval, then choose the lowest-cost one that clears your quality bar.
03We post-train the winner
We post-train the best open model to give you the best cost per outcome. Custom-built for your workflow, not a shared API.
04Production closes the loop
It runs on sovereign India servers on Dvar's platform, with production feedback continuously improving it.
CASE STUDIES
Outperform frontier models at a fraction of the cost
Route each message to the right intent.
Banking77 covers 77 customer-service intents. This 3,080-message held-out Hinglish set compares fine-tuned Kev 0.5B with the base model, Jev 1.13 and Sonnet 4.5.
Read the full case study| MODEL | ACCURACY | COST / 1,000 |
|---|---|---|
| Dvar fine-tuned Kev0.5B · 2 epochs | 88.1%Highest accuracy | ₹1.30Lowest cost |
| Sonnet 4.5Prompted | 81.2%−6.9 pp | ₹83.0263.9× higher |
| Jev 1.13Called as-is | 79.2%−8.9 pp | ₹6.905.3× higher |
| Kev, as published0.5B · no fine-tuning | 59.7%−28.4 pp | ₹1.28About the same cost |
Document extraction
12x cheaper₹0.12 vs ₹1.53
School ID card: extract 10 required fields.
Dvar vs Claude Opus 5
Inbound intent
52x cheaper₹0.03 vs ₹1.48
Customer message: identify product, urgency and PII; route it.
Dvar vs Claude Opus 5
Maker-checker
61x cheaper₹0.06 vs ₹3.50
Application checklist: decide whether to send back, and give a reason.
Dvar vs Claude Opus 5
India-hosted inference.
Deploy to fit your needs.
Start with the serving model that fits today. Change the boundary as volume, latency and governance needs evolve.
Bring us a workflow worth improving
We'll map the quality bar, model choices and path to production with your team.
Talk to a researcher