Summary
Razorpay built an AI payments foundation model trained on four billion transactions, built with NVIDIA and AWS, using ~3,000 signals per transaction to analyze payment flows across merchants, banks, instruments, and gateways. Unlike traditional ML models built for single use cases (fraud, success rate, checkout), the foundation model learns from the entire payments ecosystem and extends to new use cases. Early tests across 1.5M transactions and 51,000+ businesses showed 8-10% payment success rate improvement, 8x international card fraud detection, and 5x identification of fraudulent/disputed transactions. Razorpay positions it as "the first payment foundation model in India," arguing models built elsewhere can't be deployed due to India's unique payment architecture.
Key Points
- Trained on 4 billion transactions; built with NVIDIA and AWS
- ~3,000 signals per transaction; analyzes flows across merchants, banks, instruments, gateways
- Tests: 1.5M transactions, 51,000+ businesses; some merchants route 50-60% of volume through it
- Results: 8-10% success-rate improvement; 8x international card fraud detection; 5x fraudulent/disputed identification
- At Blinkit: 1-2 percentage point success-rate improvement
- CEO Harshil Mathur: single foundational model that works across use cases ("multidimensional")
- India-specific: UPI, cards, net banking, diverse banking ecosystem differ from other markets
- Built and operated within India; PII removed before training
- Guardrails: maintaining existing false-positive rates for fraud
- Not charging merchants separately short-to-medium term; monetization via volume + services (lending, marketing)
- Future use cases: authentication, routing, fraud, lending
- Razorpay IPO continues via confidential route