Recently I completed an AI automation project for a client who was running payment operations through WhatsApp 💬

Here's how it worked before I got involved: customers would pay via Stripe,screenshot their receipt, and send it to a WhatsApp group. Someone on the team would open Stripe, manually search for that transaction, and type a reply in the chat. They had 50+ active groups. This was happening hundreds of times a day 😅

No system to catch fake receipts. No system to catch someone sending the same screenshot twice. Just humans, doing the same lookup over and over.

So I built a bot that sits inside every group and handles it automatically 🤖

When a payment screenshot comes in, it extracts the transaction data using OCR (Tesseractl ocally, Groq Vision or Gemini as fallback for low-quality images), checks it against a duplicate store so recycled receipts don't slip through, then hits the Stripe API to verify the payment is real. The primary match key is the customer's email from the WhatsApp caption not just the transaction ID, because that alone is too easy to fake. It replies directly in the group with the verdict: ✅ verified, 🚨 fraud flagged, or 🔁 duplicate. Everything logs to Google Sheets. Anything suspicious fires a Telegram alert.

The whole thing runs as three Docker services on a DigitalOcean VPS, 24/7 ⚙️ WhatsApp bot in Node.js, OCR microservice in Python/FastAPI, and n8n handling the verification workflow. If a service crashes, it restarts automatically.

One thing I was deliberate about: the system always fails toward "unverified," never toward "verified." Low-confidence OCR goes to human review, not to an auto-approval. That decision matters a lot when you're touching payments 🔐

The client no longer has anyone manually opening Stripe. Fakes and duplicates get caught before anyone acts on them 🎯

If you are facing such problem you can contact me here and if you have some ideas how to improve this system then please discuss in comment with me !!