Stellar — Reconciliation Engine

Businesses that sell on ticket marketplaces see bank withdrawals that never quite match their sales records: merchant names are truncated ("TM *" instead of Ticketmaster), one charge can cover several orders, and settlements land days after the sale. Reconciling them by hand means hours in spreadsheets every week.

The core is a weighted confidence score computed for every charge/record pair: amount similarity (40%), date proximity (30%), merchant similarity via Levenshtein and trigram fuzzy matching with marketplace alias resolution (20%), and historical pattern analysis over the last hundred reconciliations (10%). Exact-amount matches within seven days get a confidence floor so obvious pairs never rank low.

Scores of 85+ qualify for bulk auto-reconcile; everything is an atomic transaction with full undo and an audit trail. Split matching lets one bank charge settle against up to three sales records when the amounts add up. Bank data streams in through Plaid with cursor-based pagination on a 15-minute sync, with Bill.com as a second expense source — all processed through BullMQ background jobs.

NestJS 11 on Fastify with PostgreSQL and Prisma on the backend; Next.js 16 with React 19, TanStack Query, and Zustand on the frontend. Suggestions return in under a second against thousands of unmatched transactions.

I owned the module end to end as my assignment on the team: 110+ reconciliation commits across both repos — the matching engine, Plaid sync, excluded/missing-purchases workflows, bulk unreconcile, and the reports pipeline.