Plan and method

Prove the lift on one flow, then scale it

How we would measure the business impact with Jar, what we need from Jar, and how this demo was built.

1Baseline the metric

Agree the number that matters (save rate on messaged users, reactivations, autopays kept) and capture today's messages with Dvar Eval Bench. Five-minute integration.

Week 1

2A/B test personal messages

A model trained on Jar's past campaigns and results, its tone, 8 languages and approved claims, against today's templates on a slice of users.

Weeks 2 to 6

3Scale what wins

Run it on Dvar Inference in India, Jar's data inside agreed boundaries, the same checks on every message, and a predictable cost.

On the measured lift

4Expand

Save flow and My Gold next, then SMS transaction parsing for round-ups and the WhatsApp support agent.

Next quarter

Why Jar will need this

Personal at Jar's scale needs AI

One set of models across messages, the save flow and My Gold, trained on Jar's data and run on Dvar.

Conversion. A message about the user's own streak, goal and reward is a better reason to save today than a template.
Retention and trust. Fewer lost autopays, and payouts explained before they become complaints.
It compounds. Every campaign result trains the next model. The weights are Jar's, run in India.
Cost stays under control. Small models Jar owns, not a per-message bill that grows with every user.

Built for caution

The model explains; Jar's system decides

Every number comes from Jar's systems (balance, quote, offers). A check fails any output with a number that is not in the data.
No return or price promises: checked in English, Hinglish and Indian languages.
The save flow only offers what Jar's system allows, and the cancel button always stays.
The assistant hands complaints, refunds and fraud to a person with a ticket.

For the first call

What we need from Jar

Which number would prove it: save rate on messaged users, reactivations, autopays kept? What is it today? How many push, WhatsApp and SMS messages go out a day, and how are they written today? Which rewards, coupons and streak mechanics are live, and do they expire? How many users try to pause or cancel autopay each day, and what happens on that screen today? Which tools run messaging and WhatsApp support, and is any LLM in use today? Monthly support tickets and the top five reasons (we expect withdrawal amount to be one)? What share of users use a language other than English? Who signs off on customer-facing AI: product, compliance or the founders?

How this demo was built

Method and limits

Synthetic users. Twelve made-up Jar-like users in eight languages. No Jar data is used anywhere.
Illustrative rates. Gold buy rate ₹12,450/g before 3% GST, sell rate 2.5% below the buy rate, and a 1.8% weekly change. Not live prices.
Same prompt, both sides. The frontier model and the open model on Dvar get identical instructions, facts and settings. Dvar models reason before answering; those tokens are counted in cost.
Tool use. The assistant uses a two-step JSON protocol (pick tools, then answer from the results), identical for both models.
Checks are rules, not a judge model: valid output, length per channel, language and script, numbers present in the data, banned claims, the key fact, real offers, the cancel button, the right tool.
Off the shelf today. The Dvar side is a general open model. A model post-trained on Jar's data would be smaller, faster and cheaper.
Approved claims are taken from Jar's public website: Save in 24K digital gold, 99.9% pure; Daily saving from ₹10, weekly from ₹50, monthly from ₹100; instant saving any time; Gold is stored in Brink's vaults and insured by ICICI Lombard; Vistra, the trustee, audits it; and more.