Razorpay launches Vulcan, a model trained on 4 billion payments

Razorpay has launched Vulcan, a transformer based payments model. The company says it was trained on nearly 3 trillion data points drawn from 4 billion payments, and that it reads around 3,000 signals per transaction.

Vulcan is pitched as a common intelligence layer. Routing, fraud detection, risk assessment and checkout personalisation all run off the same model, instead of the separate systems that handled each job before. NVIDIA supplied the GPUs. AWS supplied cloud infrastructure and Amazon SageMaker. Blinkit, Bachatt and redBus were named as early adopters.

What Razorpay claims

Payment success rates improved 8 to 10 percent. International card fraud detection went up 8x. Five times more fraudulent or disputed transactions were identified, without raising alert volume. Through Magic Checkout, 40 percent more shoppers were shown their preferred UPI app.

Every one of those figures is a vendor claim. Razorpay measured its own model on its own traffic and published the result. None of it has been independently verified. That is normal for a launch, and it is not a reason to dismiss the numbers. It is a reason to treat them as a starting hypothesis rather than a settled fact.

The operator point

An 8 to 10 percent swing in payment success rate is not an abstract claim. It is the one number on that list a brand can check for itself, from its own gateway logs, inside a fortnight.

Pull your authorised versus attempted transaction counts for the four weeks before Vulcan was switched on for your account, split by method: UPI intent, UPI collect, cards, netbanking. Then pull the same four weeks after. If success rate really moves 8 points on your traffic, it will show up in that table without any help from a dashboard. If it moves 1 point, that is worth knowing too, and it costs you one afternoon to find out.

One caution. Do not compare a sale week against a normal week. Traffic mix moves success rate on its own.