Razorpay launches AI model Vulcan for payments

Razorpay has launched Vulcan, which it calls India's first transformer-based AI foundation model for payments. Built with Nvidia and Amazon Web Services (AWS), the model handles payment routing, fraud detection, risk assessment and personalisation. Razorpay says Vulcan was trained on 4 billion payments and 3 trillion data points, using around 3,000 signals per transaction, and was built, trained and hosted in India. Customers including Blinkit, Bachatt and redBus have started using some capabilities in live payments, according to Entrackr.

Razorpay claims Vulcan delivers an 8-10% improvement in payment success rates, 8 times more international card fraud detection, and 5 times more fraudulent or disputed transactions identified without increasing alerts. The company is not charging merchants for Vulcan. However, medianama.com notes that Razorpay has not disclosed whether merchants can opt out of having their transaction data used for training, what legal basis it relies on for training the model (given it often acts as a data processor under the DPDP Act), or whether individual consumer data is included in the training set.

Razorpay plans to expand Vulcan into authentication and lending. The expansion comes as the company heads toward a public listing, and raises additional questions about fairness and explainability, particularly if the model is used to assess creditworthiness based on transaction patterns.