Our mission · faster recovery for banks

Resolve failed UPI transactions.
Target: under 60 seconds.

Turn a stuck payment into a verified resolution. AI agents trace the failure, apply RBI rules and carry out eligible recovery actions—with your team in control of approvals.

Under 60 seconds is our goal. Measured real-model sandbox runs: 62–85 seconds. Human review and external responses can take longer.

Evidence behind decisionsHuman approval when neededSandbox demo
Illustrated workflow · T01Live model shown in desk
Bank exception ledger

₹2,450 debited · beneficiary credit missing

  1. Master agent

    Picks the next failed payment from the ledger

  2. Investigator

    Reconciles remitter, NPCI and receiving bank records

  3. RBI Policy

    Checks the deadline, bank authority and compensation

  4. Resolver

    Resolves within limits or routes to human review

Evidence → decision

Mandatory policy gate

Actual run time and model activity are visible in the bank desk

22

seeded ledger cases

1 + 3

master + specialists

<60s

our resolution target

RBI

mandatory policy gate

Who is it for?

For the bank teams
behind every failed payment.

DisputePro gives payment operations, reconciliation and exception-handling teams one desk to investigate the payments that need intervention—and a clear record of how each case was handled.

A small percentage. An enormous workload.

UPI operates at
nearly 800 million
transactions a day.

NPCI reports 24.51 billion transactions in August 2026—about 790.6 million a day on average.

Source: NPCI · August 2026 ↗
ILLUSTRATIVE SCALE · ASSUMED FAILURE RATE
800Mdaily transactions
×
1%assumed technical failures
=
8 millionfailed attempts in a day

This is a scale illustration, not a verified network-wide failure rate. Failures that auto-reverse do not all become manual cases; the unresolved exceptions are the work this desk targets.

Why banks need agents

Every unresolved payment
has a deadline.

A backlog means more than extra work. Once an eligible failed payment crosses its RBI turnaround limit, compensation starts adding up. Agents help banks act sooner, with the policy checks and evidence still in place.

PERSON-TO-PERSON · P2PT+1

Debited. Never credited.

For a UPI funds transfer where the account is debited but the beneficiary is not credited, the beneficiary bank must reverse the funds by T+1 if it cannot credit the account.

₹100 per dayfor delay beyond T+1
PAYMENT TO MERCHANT · P2MT+5

Debited. Merchant unconfirmed.

For a UPI merchant payment where the account is debited but confirmation does not reach the merchant, the auto-reversal deadline is T+5.

₹100 per dayfor delay beyond T+5

T is the calendar date of the transaction. Eligible compensation must be credited automatically, without waiting for a customer complaint. This is customer compensation under RBI’s framework, not a flat fine on every failed attempt. Read the RBI circular · UPI rows 4(a) and 4(b) ↗

What delay can cost

Small backlogs.
Real money.

See how compensation adds up when eligible cases remain unresolved beyond their deadline.

Illustration only · not a forecast of your bank’s liability
Illustrative compensation₹1,00,0001,000 cases × 1 day × ₹100

Speed is useful. Governed speed is the product. Agents collect the evidence and prepare the remedy. Mandatory policy checks, fresh ledger verification and approval limits decide what can execute.

Meet the agents

Meet your payment
recovery team.

One master agent coordinates three specialists. Each has a defined job, visible model activity and source-backed findings. Your analysts handle the cases that need judgment.

Master agent

Picks the next failed payment from the ledger and coordinates three specialists, one case at a time.

Investigator

Checks payment records across the bank, payment network and receiving bank.

RBI Policy

A mandatory gate checks T+1 or T+5, bank authority and compensation before financial action.

Resolver

Assesses risk and recommends the next action, with evidence for the reviewer.

Checks before action

The execution engine checks the payment and ledger again before acting.

Approval for sensitive cases

Fraud cases and payouts over ₹10,000 wait for a human decision.

An auditable case receipt

The desk preserves the ruling, executed actions and references for bank analysts.

Built at neatHack

An agentic AI solution.
Built at neatHack.

Thank you to the sponsors for helping us turn a difficult banking problem into a better solution. Your tools made the agents observable, the build traceable and the business plan more grounded.

Agent observability

neatlogs

Connected

See what each agent did, which tools it called and where a run slowed down.

Build history

Entire

Connected

Follow the AI sessions, checkpoints and commits that shaped the product.

Business planning

cfo.ai

Model verified

An editable 12-month model with pricing, delivery costs, cash flow and three scenarios.

Inspect verified work ↗

Company intelligence

Crustdata

API verified

Live company research feeds the business plan. Ambiguous matches stay visible for review.

Inspect verified work ↗

Build evidence · updated 11 October 2026

See what works.
Inspect what proves it.

The latest bank desk and a recorded resolution, with the measured numbers behind them. Screenshots open here, so you can inspect the details without losing your place.

21

backend tests passed

Includes approvals, duplicate protection and changed-evidence holds.

62.1s

measured outward refund

A real four-agent sandbox run; under 60 seconds remains the goal.

64.1s

measured inward return

Funds returned from suspense with seven source calls.

Project evidence · 1 / 2

The product · A bank desk built around agents

The product · A bank desk built around agents

The failed-payment ledger, the payment chain, a master agent with three specialists, and RBI governance with the outcome side by side.

Model runs use sandbox bank and network adapters. Business forecasts are assumptions, not measured revenue. Sponsor research and build history are separate from transaction evidence.

Try the full journey

Let the agents take
the next case.

Open the bank desk. Follow a failed payment from investigation to policy check, action and receipt. See the work, not just the result.