Their world
Razorpay has redrawn its architecture around agentic payments, where an agent can discover, decide, and complete a transaction on its own.
The hard problem is no longer shipping the first agent. It is running many of them in production, where every action a model takes in the payment path has to be measurable, reversible, and reviewable.
Their pressure
You are launching for it and hiring for it — and now you own the operational surface it creates.
- Architecture Razorpay revamped its payments architecture around agentic AI on the Claude Agent SDK at FTX 2026.
- Launch Razorpay Agent Studio and an Agentic Experience Platform shipped at FTX'26 for onboarding, integration, and payment operations.
- Hiring Open roles for AI Builders in multi-agent systems, RAG, and production AI pipelines.
The models are the easy part. The evals, guardrails, and rollback paths that make an agent safe to run at Razorpay's transaction volume are the work — and they are what slows a demo on its way to production.
The bridge
Meridian AI embeds with your team and builds the production layer around your agents.
We are a boutique applied-AI studio that ships production-grade LLM and agent systems — on your data, with human-in-the-loop controls — in six weeks, not six quarters. For Razorpay that means agents wired into your current services, with evals and guardrails in CI, so a new agentic payment flow reaches production with its checks already attached.
One proof
We ship on your existing stack — retrieval over your own data, function-calling agents wired into your current services, and evals plus guardrails in CI. No model training, no rip-and-replace, no new infrastructure to run.
One working session
Map one Razorpay agent flow from request to reviewable decision.
Twenty minutes on one live agentic-payment workflow: the task boundary, the evals, the guardrail, and the human-review path.
See the 20-minute teardown for Razorpay