Building a Bank Reconciliation Engine with Fuzzy Matching

How I built a progressive fuzzy-match reconciliation engine across 5 bank accounts — tier-priority confidence scoring, review queues for ambiguous matches, and discrepancies that surface before month-end, not after.

The problem

Traditional bank rec tools assume perfect data. Real life is messy — memos change, amounts split across multiple posts, dates drift by a day, and descriptions shift between banks. A pure exact-match engine breaks on the first edge case.

What you'll build

• A multi-tier match engine (exact → date-adjusted → fuzzy-memo → amount-split)
• Confidence scoring so you know which matches to trust
• A review queue that surfaces only what a human needs to look at
• Audit logs that make every match reversible

Prerequisites

• Python 3.11+ with pandas and rapidfuzz
• Sample bank statements and a GL/AP export
• Basic comfort with SQL and dataframe joins

1. Designing the match pipeline

Coming soon — why progressive matching beats a single pass, and the tier order I use.

2. Writing the match tiers

Coming soon — exact, date-window, fuzzy-memo, and amount-split matchers — with code.

3. Confidence scoring

Coming soon — how I weight each signal and set a threshold for "auto" vs. "review."

4. The review queue

Coming soon — the minimal UI I render as HTML + the audit log that survives re-runs.

Wrap-up

Coming soon — what went wrong on the first version and how I fixed it.