Aleph Hackathon 2026

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. If that's not acceptable, use the local-folder mode only.

Quick start

python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

npm install -g @qvac/sdk@0.17.1
export QVAC_SDK_DIR="$(npm root -g)/@qvac/sdk"        # Windows: $env:QVAC_SDK_DIR = "$env:APPDATA\npm\node_modules\@qvac\sdk"

python scripts/generate_sample_data.py   # demo dataset: 21 receipts + bank_statement.csv
python main.py                           # runs OCR + reconciliation, writes reports/

Exits with code 2 if a CRITICAL discrepancy is found (useful for cron/CI). Key CLI flags: --receipts-dir, --bank-csv, --whatsapp-month, --date-tolerance-days, --amount-tolerance, --merchant-threshold, --large-unmatched-amount.

Other ways to run it

• python main_web.py — local web UI (http://127.0.0.1:8080) for picking a receipt source and bank CSV without touching the CLI.
• python main_whatsapp.py — webhook server for real WhatsApp Business ingestion. Try python scripts/simulate_whatsapp_traffic.py first to see the flow without a Meta account.

What the matcher flags

For each receipt: find bank transactions within a date window, fuzzy-match the merchant name, and compare amounts. Outcomes: clean match, AMOUNT_MISMATCH, MISSING_IN_BANK, DUPLICATE_RECEIPT (resubmitted receipt), or UNACCOUNTED_CHARGE (bank activity with no matching receipt). The matcher never guesses — low-confidence OCR reads get flagged for manual review instead of forced into a match.

Project structure

main.py / main_whatsapp.py / main_web.py   # entry points
src/reconciliation_agent/
├── ocr_engine.py       # QVAC integration (the only inference layer)
├── extractor.py        # OCR text -> structured Receipt
├── bank_loader.py       # bank CSV -> normalized DataFrame
├── matcher.py           # reconciliation business logic
├── report.py / report_html.py
├── webapp.py            # local web UI
└── whatsapp/            # WhatsApp Business API ingestion
scripts/                 # sample data generators, WhatsApp traffic simulator
data/                    # sample receipts + bank statements (US and AR formats)
tests/

Tests

pip install -r requirements-dev.txt
pytest

QVAC and the Meta API are mocked — no credentials or network needed.

Builders Contact

• Felipe Bridge — https://www.linkedin.com/in/felipebridge/
• Emiliano Lescuras — https://www.linkedin.com/in/emilianolescuras/
• Matias Bellido — https://www.linkedin.com/in/matiasbellido