Every payment tied to its order. Automatically, and provably.
A deterministic engine links each payment to its order across four lists that never agree - by reference, by amount and time, and by structured pattern. It never guesses silently.
Matching is where correctness lives, so it is deliberately deterministic and auditable - rules and passes decide, the AI never does. Payments run through ordered passes, strongest evidence first, and the first confident hit wins.
Anything the engine can't match with confidence doesn't get forced into your books. It becomes a typed exception with the evidence attached, so a human decides - and teaches the engine a rule for next time.
What you get
- Exact-reference matching - order id or gateway ref on both sides
- Amount + time-window + provider matching with tie-breaks
- Structured patterns - split payments, bundled transfers, part-payments
- Zero tolerance on order amounts - a rupee off is a mismatch
- Every match evidence-linked, with the confidence and rule recorded
- Merchant-scoped rules learned from human resolutions
The matching engine, in detail.
Passes, strongest evidence first
Pass 1 matches on an exact reference - an order id or gateway ref present on both sides - and auto-commits at full confidence. Pass 2 matches on an equal amount inside a time window with the right provider, breaking ties on the closest timestamp. Pass 3 takes on the messy shapes. Whatever survives all passes becomes a typed exception rather than a forced entry.
The shapes real money makes
Real payments rarely map one-to-one. One order is settled in two wallet payments; one Fonepay transfer covers a stack of invoices; a payment arrives as an order minus a known discount. Pass 3 recognises these structured shapes and ties the parts back to the whole, so a split or bundled payment is reconciled as a unit instead of landing in the exception queue.
Tolerance is zero where it matters
For tying a payment to an order, a rupee off is a mismatch - full stop. Small fee-rounding differences are auto-explained with the evidence that justifies them, but anything larger is surfaced rather than absorbed. A near-match that would quietly paper over a shortfall is exactly the thing that gets flagged, because money is not a place for 'close enough'.
It learns - on a leash
When a person resolves an exception and ticks 'remember this', the engine records a rule scoped to your organization alone. Rules act only as evidence for a pass, never as a hidden override of the state machine, and any rule a human later overturns is auto-disabled. The audit trail keeps every rule and every match it influenced, so accuracy compounds without you losing control.
Deterministic by design, AI on the side
Correctness lives in the passes, so the passes alone decide - the AI never writes to the ledger. On an exception it may propose a one-line hypothesis with a confidence score, and it can explain a finished match in plain words, but the number and the decision come from deterministic rules. There is no code path from the model to your books.
FAQ
The matching engine,
explained
Common questions about the matching engine. Want it run on your own data first? The concierge reconciliation is free.
On real merchant data the target is 95%+ auto-matched, with 100% of the rest surfaced as typed exceptions and zero silent discards.
Related products
Get your free reconciliation.
Send us last month's eSewa report, Khalti export, bank statement and orders sheet. Within 48 hours we send back a one-page close - every payment matched, every fee computed, every settlement decomposed.
Files only · private upload link · nothing to install