Summary
Razorpay built "Vulcan," an AI payments foundation model trained on four billion transactions, to improve payment success, fraud detection, and checkout personalisation. Built with NVIDIA and AWS using ~3,000 signals per transaction, it analyses payment flows across merchants, banks, instruments, and gateways. Unlike traditional ML models built for individual use cases, Vulcan is a single multidimensional foundation model that learns from the entire payments ecosystem. It delivered an 8-10% improvement in success rates, 8x international card fraud detection, and 5x detection of fraudulent/disputed transactions in tests across 1.5M transactions and 51,000+ businesses.
Key Points
- "Vulcan": trained on 4 billion transactions; ~3,000 signals per transaction
- Built with NVIDIA and AWS; operated within India; PII removed before training
- Tests: 1.5M transactions, 51,000+ businesses
- 8-10% improvement in payment success rates
- 8x increase in international card fraud detection
- 5x increase in detection of fraudulent/disputed transactions
- Some merchants route 50-60% of volumes through it; Blinkit saw 1-2pp success improvement
- Single multidimensional foundation model (vs siloed use-case ML models)
- India-specific: UPI, cards, net banking, diverse banking ecosystem — foreign models not directly deployable
- First payments foundation model in India; "globally, very few companies have built"
- Not charging merchants separately; monetising via volumes + lending/marketing applications
- Guardrails: maintains existing false-positive rates for fraud detection
- CEO Harshil Mathur: combine AI with financial services; potential across authentication, routing, fraud, lending
- IPO continues via confidential route, "on track"