Code decides, not the model.
The case stays open until the system shows the fix.

Sahkarmi is a tested rule layer for a merchant-desk AI agent. It is a hackathon prototype on a simulated back office, and the ask is simple: a shadow run on real cases that acts on nothing.

Pick a case

Recordings of the real agent on a simulated back office. Left: the shop's phone. Right: what Sahkarmi did.

Shop's phone

SAHKARMI · MERCHANT DESK

Hinglish demo

The whole idea in 2:43.

Recorded on the prototype, with what works, what does not, and what we are asking for.

The problem

The shop asks why the money is short. The ticket closes anyway.

"Kirana sales don't have invoices, so there's nothing to upload."

A merchant on Reddit, 24 Oct 2026: 11 disputes, ₹960, taken from the next settlement. Sahkarmi does not fully solve this case yet. Read the post

UPI chargebacks accepted or deemed accepted, Sep 2026

Paytm's line56%

PhonePe's line11%

Our maths on NPCI's table. Reading YES BANK PTY as Paytm's line is our inference, and NPCI gives no reason. Other large lines sit between 11% and 35%. Chargebacks are a severe edge case for a small shop, not its daily pain.

Who else is doing this

The idea is not new, and we say so.

Paytm owns the shops, the data and the reach, and it could build all of this. What we bring is a stated rule, tested in the open: code decides, and a case closes only after a system check.

What they show in public	Paytm	BharatPe	Razorpay	Sahkarmi

From public sources, Oct 2026. "Not shown" means not found in public, not that it does not exist. Sahkarmi's column is a prototype on simulated systems.

The four questions

Would it work at Paytm's size?

Z-REPORT · TEST RUNS31 OCT 2026

What is not real yet

Tested on a simulation. Nobody has used it for real.

Our own audit found four holes on 31 Oct: one Confirm tap could pay repeated refunds, a reworded reason skipped the once-only rule, "ok thanks" counted as consent, and a case could close with a problem still open. Those four cases no longer run, and each has a test. Tying consent to a specific offer is still open.

The back office is a test copy. No real Paytm API has been called.
No real merchant has used it, and no real money has moved.
Soundbox has no public API, so its steps are simulated.
A shop with no invoice on a "goods not received" claim is still an open gap.
Customer details are not yet masked before they reach the model, and the demo has no login.