The model proposes. Policy decides.
Every action is checked against a written policy using data re-read from the source systems, not what the model said. Refunds, blocks and disputes need a single-use approval tied to the exact amount.
Paytm Build for India AI Hackathon · Track 3
An autonomous AI teammate for customer support. It investigates cases across systems, acts under written policy, verifies its own work and resolves the customer's issue. Anything risky goes to a person, with the whole case file.
Built and demonstrated on payment support workflows
Or talk to it, in Hindi or English. Voice by Sarvam AI.
Code runs the loop. The model is only asked for small, checked answers along the way.
Reads the complaint in plain words, Hindi or English.
Pulls the customer's records with read-only tools.
Proposes a fix under a written, versioned policy.
Runs the action only if the policy engine authorizes it.
Reads the systems back to prove it actually happened.
Tells the customer what was done, with references.
Or escalate: when the policy says a person must decide (fraud, big amounts, a failed check), Kriya prepares the handoff, and carries out whatever the person approves.
Every action is checked against a written policy using data re-read from the source systems, not what the model said. Refunds, blocks and disputes need a single-use approval tied to the exact amount.
After acting, Kriya reads the ledger back. If a refund fails, it does not tell the customer it worked. It hands the case to a person instead.
The live trace shows each tool call, rule and check. The audit log is append-only and hash-chained, so any tampering shows.
Customer portal
Read what happened to them, then write to support in your own words, or call. A failed payment, a double charge, payments nobody made, a missing settlement.
Open the portalOps console
Watch cases move live, open a case file, approve what Kriya escalated, change its autonomy limits and see its behaviour change.
Open the console