Razorpay (@Razorpay) / X

Razorpay

7,433 posts

Razorpay

@Razorpay

Backing India's Boldest Founders. Join the movement 👇

India

Joined July 2014

Affiliates Replies Media

Razorpay
@Razorpay
15h
Deploying voice agents for cart recovery isn't like deploying traditional software. The non-deterministic nature of LLMs introduces risks like hallucinations and policy violations. That's why every Agent Studio agent first runs in a no-customer-contact state we call Shadow Mode.
Razorpay
@Razorpay
15h
Replying to @Razorpay
We evaluate these execution traces against strict guardrails for safety, compliance, and fabrication. For example, during early evaluations of our cart recovery agent, Shadow Mode revealed a 30% fabrication rate where the model invented unapproved discount codes.
Razorpay
@Razorpay
15h
By isolating this behavior, we were able to debug the prompt harness and agent logic, reducing the fabrication rate to <1%. Shadow Mode gives us statistically significant data to guarantee an agent's reliability before it ever enters production. In other words, it ensures our AI
Razorpay
@Razorpay
16h
Payments infrastructure is never really finished. Every year brings new ideas around security, identity, fraud prevention and user experience. Looking forward to seeing what builders across India create through NPCI's Token Nxt initiative. Which part of the payments stack do
Razorpay
@Razorpay
19h
Thousands of international Shopify stores have Indian customers browsing their products right now. Most of them can't convert those visitors. Because India's payment methods, compliance rules, and checkout expectations are completely different. We just fixed that- with one
Razorpay
@Razorpay
19h
Replying to @Razorpay
"But, what about compliance?" India requires buyer name + address on every cross-border transaction. PAN and DOB when AML rules trigger. Razorpay's checkout handles all of it automatically - collects what's needed, when it's needed. Compliance is built into the checkout, so
Razorpay
@Razorpay
19h
Collect in INR. Settle in USD, EUR, SGD, GBP, or 30+ currencies - directly to your overseas bank account. Your Shopify store just became an India-ready business. Install the Plugin and get started today →
apps.shopify.com
Razorpay Trusted Business - Improve sales by displaying trust widget on your product pages |...
A significant number of shoppers' drop-offs happen because of lack of trust in a business. Razorpay Trusted Business Widget helps your business bui...
Razorpay
@Razorpay
Jul 19
When anyone can build an application, the bigger question is: what should you build next? Watch Vishal Virani, Co-founder of @rocketdotnew, explain why the next generation of AI tools will move beyond generating code to tracking performance, understanding user behaviour, and
00:00
Razorpay
@Razorpay
Jul 18
A historic day for India's private space industry. 🇮🇳🚀 Congratulations to @SkyrootA on the successful launch of Vikram-1, India's first privately developed orbital rocket. Proud to see Indian companies continue pushing the boundaries of what's possible.
Razorpay
@Razorpay
Jul 17
Here's why Razorpay's Agent Platform is built around a single-agent architecture. (Part 2) 🧵
Razorpay
@Razorpay
Jul 17
Replying to @Razorpay
3/4 Every new capability is an assembly of primitives, not a bespoke subsystem. A next-best-action suggestion is just a pre-conditioned entry point into a tool the agent already has. A multi-step write is a skill plus the workflow store plus the same safety gate. The platform
Razorpay
@Razorpay
Jul 17
4 /4 The core is surface-agnostic by design. A new front door plugs in with just an ingress adapter, an auth resolver, and a render node, the brain underneath never changes.
Razorpay
@Razorpay
Jul 17
Most AI platforms scale by adding more agents. We went the other way. Here's why Razorpay's Agent Platform is built around a single-agent architecture. (Part 1) 🧵
Razorpay
@Razorpay
Jul 17
Replying to @Razorpay
2/3 It's a single thin orchestrator that composes of four primitives: tools, skills, memory, and knowledge. Domains don't become agents, they become skills and tools the one agent composes. It finds the right one on demand instead of routing between domain agents.
Razorpay
@Razorpay
Jul 17
3/3 The core trick is how it finds what to do. Loading every tool definition into context stops working once you're in the hundreds, selection accuracy falls apart. So the agent keeps the catalog out of context and pulls only what the current turn needs. To be continued.