Razorpay Unveils Vulcan AI Model to Optimize Payment Routing and Fraud Detection

The model, named Vulcan, was trained on nearly three trillion data points across four billion payments and is designed to consolidate several payment-intelligence functions under one shared layer. For developers and payment platforms, the important shift is architectural: routing, risk assessment and fraud controls are being handled by a common model rather than isolated machine-learning systems.

A shared model for transaction decisions

Vulcan analyses about 3,000 signals per transaction, according to the report. Those signals are used to select the payment route most likely to succeed, detect fraud across merchants, assess return-to-origin risks for cash-on-delivery orders and personalise the checkout experience.

The model was trained and operated on NVIDIA GPUs, while AWS supported development and deployment through Amazon SageMaker. The deployment therefore combines a specialised compute layer with managed model infrastructure, rather than presenting Vulcan as a standalone application model exposed directly to end users.

Razorpay’s stated objective is to create a feedback loop across its payments network: transaction outcomes become additional information for later decisions. In practical terms, that could allow routing and risk systems to be updated through a common intelligence layer instead of requiring each subsystem to be retrained and tuned independently.

The available reporting does not provide Vulcan’s parameter count, inference latency, model architecture or quantization strategy. Those omissions matter for engineering teams evaluating whether the system represents a foundation-model change in the technical sense or a broader consolidation of existing payment models behind a common platform.

Reported performance in live payments

Early components of Vulcan are already running on Razorpay’s network, with Blinkit, Bachatt and redBus among the customers using them in live payment environments, Fintech Singapore reported.

Razorpay reported an 8% to 10% improvement in payment success rates. The system also detected and stopped eight times more international card fraud and identified five times more fraudulent or disputed transactions without increasing the number of alerts, according to the same report.

Those figures describe operational outcomes rather than standardised model benchmarks. They indicate where Razorpay says the system has produced value—transaction completion and fraud detection—but do not establish how the model compares with other payment-routing systems under identical workloads, latency targets or fraud-labeling conditions.

The checkout layer is another deployment point. Razorpay said Magic Checkout showed 40% more shoppers their preferred UPI application, helping merchants complete between 100,000 and 200,000 additional purchases each month. The reported result connects model output to payment conversion, but the evidence available here does not specify the baseline period or the precise contribution of Vulcan relative to other changes in the checkout flow.

Why the launch matters for AI infrastructure

Payment routing is a high-frequency inference workload in which model quality is constrained by more than classification accuracy. Decisions must be made across multiple signals while transaction paths, fraud patterns and merchant behaviour change over time. A shared model can reduce duplicated feature pipelines and make cross-system information available to routing, fraud and risk components—but it also creates a larger dependency on data governance, evaluation and failure isolation.

Razorpay said it eventually aims to use Vulcan across authentication, routing, fraud detection and lending. That would extend the model from transaction-time decisions into a wider financial-services stack, increasing the importance of monitoring how errors propagate between functions.

For the AI industry, the launch is a concrete example of foundation-model positioning outside general-purpose language systems. The value proposition is not conversational capability or a published benchmark score; it is the ability to process a large, domain-specific transaction history and apply a shared inference layer to operational decisions. The next technical details to watch are latency under production load, model refresh cycles, the separation of fraud and routing objectives, and whether the reported gains persist as deployment expands across Razorpay’s network.