A payment-run validator that clears the approver's anomaly check in minutes, not hours

Context

Every payment run, each approver has to satisfy themselves that nothing in the file is wrong before they release money — a changed bank account, a step-up against the vendor's historical payments, a duplicate, an amount that doesn't fit the pattern. Done properly, that anomaly check is eyeballing a payment file against history line by line. It takes 30–40 minutes per approver, every run, and it's exactly the kind of slow manual control that either becomes a bottleneck or gets rubber-stamped.

Challenge

Speed it up without weakening it — and do it on data you cannot hand to a model. Bank statements and account numbers can't be pasted into an LLM; the validation also has to be *deterministic and explainable* (an approver releasing money needs the same answer every time, not a probabilistic one). And whatever clears the run has to stand up later: an auditor will ask which exact file was approved, by whom, and when.

What I built

A local Python desktop app with a GUI an analyst can run without training. The workflow is two clicks: select the payment-run file, select the history to check it against — and the full reconciliation runs on the machine, nothing leaving it.

• Deterministic rule engine — flags bank-account changes against history, step-ups versus a vendor's historical payment pattern, duplicates, and amount anomalies. Same input, same result, every time — and every flag is explainable, not a black-box score.
• Local and private by design — runs entirely on the desktop. Sensitive bank data is never sent to any model or external service. This was a deliberate choice: the job called for determinism and confidentiality, so it gets a tool, not an AI.
• Tamper-evident audit trail — the app records the files used, the username, and the timestamp, and stamps the SHA-256 hash of each input file, so there's no ambiguity later about exactly which version was validated.
• Beautiful PDF audit report — a clean report generated alongside each payment run, carrying the full audit trail and the anomaly findings, ready to attach to the approval.

Outcome

The per-approval anomaly check dropped from 30–40 minutes a run to about 5 minutes — and the control got *stronger*, not weaker: every release now carries a hash-verified, timestamped audit report instead of a tired analyst's word that they looked. The bottleneck in the payment run is gone, and the audit question — "which file did we approve, and who approved it?" — answers itself.