TasdeeqPay

Pakistani merchants reconcile manual bank transfers by hand: a customer pays, screenshots the receipt, and someone eyeballs it against the order and the bank alert. It's slow, and the errors are asymmetric — a wrong 'paid' means you shipped for free.

TasdeeqPay automates that reconciliation with a deterministic engine and a hardened mobile pipe, and it's tuned so the expensive mistake can't happen. It's a working prototype of the core; productionisation is in progress.

Three sources, one agreement

The engine trusts no single input. It cross-matches three independent signals — the order (what's owed, to whom), the customer's payment screenshot read via OCR, and the merchant's own bank alert. It confirms only when all three agree. Partial agreement isn't a 'probably'; it's a hold for human review.

Bias toward the cheaper error

Every classifier makes two kinds of mistake, and in payments they aren't equal. A false positive — confirming a payment that didn't clear — costs the merchant the value of the goods. A false negative — missing a real payment — costs one manual check. So the system is tuned to eliminate the first at the expense of tolerating the second.

Let the model read, not a brittle parser

Every bank formats statements differently. A hardcoded parser per bank breaks the moment a layout changes; an LLM extracts the fields and generalises to new formats instead of throwing. The determinism lives in the matching logic; the flexibility lives in the reading.

Getting the alert off the phone

Bank alerts arrive on an Android phone, and Android OEMs aggressively kill background apps. The companion app uses an outbox pattern hardened against exactly that: the alert is persisted the instant it arrives and forwarded reliably, not best-effort. Correctness at the core is worthless if the pipe feeding it drops messages.