Vulcan: Razorpay’s Innovative AI Foundation Model for Payments

Fintech unicorn Razorpay has introduced Vulcan, an AI foundation model specifically tailored for optimising payment processes, aided by technology from NVIDIA and AWS. This Bengaluru-based firm has reported that the transformer-based model has been trained using nearly 3 trillion data points derived from 4 billion payment transactions, employing around 3,000 signals per transaction to inform payment-related choices.

The training and operation of this model were facilitated by NVIDIA GPUs, while the developmental phase, including training and deployment, was supported by AWS cloud infrastructure and Amazon SageMaker. Razorpay stated that Vulcan functions as a unified intelligence layer for various payment operations, which include routing, fraud detection, risk assessment, and checkout personalisation, thereby replacing the need for separate models for distinct functions.

The components of this model have already been tested on Razorpay’s network prior to the official launch. Several customers, such as Blinkit, Bachatt, and redBus, have begun utilising some of these features in real-time payment scenarios. Razorpay claims that this technology has enhanced payment success rates by 8-10% and facilitated the detection and prevention of 8 times more international card fraud. It has also successfully identified five times more fraudulent or disputed transactions without escalating the number of alerts.

Through Magic Checkout, the company indicated that 40% more shoppers now see their preferred UPI application, contributing to an additional 1-2 lakh transactions each month. The model leverages signals from merchants, payment instruments, issuers, and gateways to evaluate payment pathways and determine the one with the highest likelihood of success. Furthermore, it can discern fraud patterns among merchants, evaluate risks linked to cash-on-delivery orders, and suggest payment methods during the checkout stage.

Razorpay has declared that Vulcan has been built from the ground up, possessing proprietary architecture and training data. An internal study conducted by the company, encompassing 1.5 million shoppers and over 51,000 businesses, discovered payment friction issues such as failed transactions, delays, and customer drop-offs in both metro and smaller markets. Razorpay intends to broaden the application of this model to encompass other sectors, including authentication and lending, as it aims to implement a single AI layer across a wider array of payment decisions. The launch coincides with predictions that India’s digital ecommerce market will soar to $350 billion by 2030, as estimated by Razorpay.