AI Financial Reconciliation Agent

An agent for small accounting firms: it reconciles client receipts against bank statements and flags billing errors or fraud. OCR and matching run 100% locally via QVAC — financial data never leaves the machine.

Built for the Tether QVAC Track: Local agents for operations work.

The problem

Small accounting firms reconcile client receipts against bank statements by hand every month — hunting for duplicate charges, mismatched amounts, and unexplained bank fees. The data is sensitive (client financials, covered by professional confidentiality), so it can't be sent to a cloud LLM API. This agent automates the reconciliation while keeping all inference on-device.

How it works

Receipts come in two ways — a local folder, or photos clients send over WhatsApp — and converge on the same pipeline:

receipts (folder or WhatsApp) → OCR (QVAC, local) → extractor → matcher (vs. bank CSV) → report (txt/html)

• Local, always: OCR (QVAC vision model on-device), matching (pandas + RapidFuzz), report generation.
• Cloud-mediated: only the WhatsApp intake path — attachments transit through Meta's WhatsApp Business API before reaching the local webhook. No other module makes network calls.

What the matcher flags

For each receipt: it finds bank transactions within a date window, fuzzy-matches the merchant name, and compares amounts. Outcomes: clean match, AMOUNT_MISMATCH, MISSING_IN_BANK, DUPLICATE_RECEIPT (resubmitted receipt), or UNACCOUNTED_CHARGE (bank activity with no matching receipt). Low-confidence OCR reads get flagged for manual review instead of forced into a match.

Builders

• Felipe Bridge
• Emiliano Lescuras
• Matias Bellido