Razorpay has introduced Vulcan, described as India’s transformer-based artificial intelligence foundation model purpose-built for payments. Developed in collaboration with NVIDIA (NASDAQ:NVDA) and Amazon Web Services, the system aims to enhance the reliability, security, and efficiency of digital transactions across the country’s complex payments landscape.
Unlike conventional machine learning tools that tackle isolated challenges such as fraud checks or success-rate optimization, Vulcan functions as a unified intelligence layer.
It processes roughly 3,000 signals from each transaction and has been trained on nearly three trillion data points drawn from four billion payments.
This scale allows the model to recognize intricate patterns in how money moves through UPI, cards, net banking, wallets, and cash-on-delivery options involving hundreds of banks and networks.
The architecture draws on transformer technology similar to that used in large language models, yet it is specifically engineered for payment data rather than text.
Razorpay developed both the model design and its training dataset in-house.
NVIDIA supplied the high-performance GPU clusters needed for training and inference, while AWS provided scalable cloud infrastructure, including tools such as Amazon SageMaker, to support development, deployment, and real-time decision-making at enterprise volumes.
Early deployments have already delivered measurable gains.
In testing involving more than 1.5 million transactions across over 50,000 merchants, including platforms such as Blinkit, Bachatt, and redBus, the model produced an 8 to 10 percent lift in payment success rates.
It also detected and blocked eight times more international card fraud and identified five times more fraudulent or disputed transactions without raising the volume of alerts sent to businesses.
On the checkout side, predictive personalization helped 40 percent more shoppers see their preferred UPI application, contributing to an estimated 100,000 to 200,000 additional completed purchases each month.
These improvements address persistent friction points that affect consumers in both large cities and smaller towns.
Failed transactions, drop-offs, and delays remain common hurdles even as India’s digital payments infrastructure handles enormous volumes.
By continuously learning from every new payment, Vulcan is designed to compound its understanding over time, refining routing decisions in milliseconds, flagging network-level fraud patterns invisible to single-merchant systems, assessing risk for cash-on-delivery orders, and personalizing the checkout experience.
Company leaders have emphasized that no ready-made foundation model suited to India’s payment architecture previously existed.
Off-the-shelf language models excel at conversation or code but cannot capture the relational dynamics of transaction graphs involving multiple banks, instruments, and gateways.
Vulcan therefore represents a specialized approach tailored to local conditions while positioning the fintech firm for further applications in authentication, lending, and broader financial services.
As India’s e-commerce market continues its projected expansion, tools that raise success rates and reduce losses become increasingly valuable. Razorpay positions Vulcan as foundational infrastructure that grows smarter with each transaction, supporting more seamless digital commerce at national scale.