AI Engineer (Platform)

AI Engineer (Platform)

AI Engineer (Infrastructure)

Razorpay’s infrastructure runs millions of payment transactions a day across payment routing, settlement, reconciliation, merchant onboarding, fraud, and lending. That infrastructure is large, always-on, and expensive to operate the way most companies operate it — with war rooms, runbooks, and humans paging each other at 3 AM. We don’t think that’s the ceiling.
This role exists to take the next layer off human shoulders. You’ll sit inside the Infrastructure team and build the AI systems that run, heal, and optimize the platform themselves — agents that triage incidents, pipelines that self-remediate, copilots that compress a six-hour Sev2 into thirty minutes, automation that catches regressions before they ship. Not chatbots bolted onto a dashboard. Infrastructure that gets smarter every quarter. You’ll report into the Infrastructure team lead. You’ll own automation and AI initiatives end-to- end — from the problem statement to the system that runs in production. You won’t have a JD; you’ll have a mandate.

So you know what you’re automating. Razorpay’s infra stack is, in broad strokes:
• Cloud. multi-cloud — AWS as the primary, GCP for specific workloads, with ongoing consolidation.
• Orchestration. thousands of nodes across EKS clusters, with service mesh, autoscaling, and progressive delivery.
• Provisioning. Terraform + Atlantis for IaC, Argo CD for GitOps, internal platforms for one-click provisioning.
• Observability. Prometheus, Thanos, Loki, Grafana, OpenTelemetry — plus a heavy internal investment in structured logging and SLO-based alerting.
• Incident response. PagerDuty-driven on-call, runbooks in Notion, and a strong — and very manual — incident response culture we want to automate aggressively.
• Data plane. high-throughput Kafka, self-managed and managed Postgres/MySQL, Redis, and a growing footprint of vector stores for the AI work itself.
You don’t need to have operated all of this yourself. You do need to be comfortable enough in this world to know which parts are painful, which parts are load-bearing, and where an agent can actually replace a human.

What you’ll actually do :
• Take an infrastructure problem from “we should automate this” to “it’s running in production, here’s the toil it removed” — in weeks, not quarters.
• Build agents that triage alerts, correlate signals across services, draft or execute remediations, and escalate to humans only when it actually matters.
• Ship LLM-powered copilots for the on-call engineer — surfacing “this has happened before, here’s what fixed it” at 3 AM instead of a blank Slack thread.
• Automate the runbook. If a human is copy-pasting steps out of a Notion doc, that’s a target. Turn runbooks into agentic workflows that run themselves.
• Build eval harnesses for infra agents — because “the AI said to restart the cluster” is only acceptable if you can show, with evidence, that it was right.
• Work alongside SREs, platform engineers, and devs — not as a service desk, but as the person who ships the thing that makes the service desk unnecessary.
• Run experiments. Have strong opinions about which models, which tools, and which agent architectures actually work for infrastructure — and prove them with output.

What we’re looking for: 

• Background. 2–6 years of experience. Top engineering or CS school in India or abroad. Or — you’re currently at one of those schools and have decided dropping out and building is the better trade. Both paths land here.
• AI-first by default. Not “I use ChatGPT sometimes.” Your daily workflow assumes AI is in the loop on every task that can be improved by it. Cursor, Claude Code, custom agents, MCPs — pick your stack and own it.
• Open-source model fluency. You’ve stress-tested the lineup — Llama, Mistral, Qwen, DeepSeek, GLM, Kimi, and whatever’s on the leaderboard this week. You have opinions on which one wins where, what they’re bad at, and where the cost / latency / quality frontier sits today.
• Infrastructure literacy. You’ve touched real infrastructure — cloud, Kubernetes, Terraform, CI/CD, observability — enough to know what a Sev2 feels like and why nobody wants to page at 3 AM. You don’t need to have run a 200-person SRE org. You do need to be able to read a Prometheus query and a Terraform plan without googling every line.
• Builder, not talker. Side projects, OSS contributions, internal tools at your last job, a hackathon weapon — something we can look at and say “that’s a person who ships.”
• Single-person army energy. Ambiguity, multi-hat, no playbook — energising rather than draining. You know good design, good UX, good code, and good prompts when you see them. You can articulate why one approach is better than another in a way that holds up under scrutiny.
• Eval mindset. You don’t just “vibe-check” your AI work. You build the smallest possible eval before shipping — and have an opinion about why most AI products fail without one.

What makes you stand out
• You’ve built agents that orchestrate other agents, not just single-shot LLM calls.
• You’ve shipped an AI-driven automation into a real on-call or ops workflow — and it’s still running without you.
• You’ve fine-tuned a model — even if just for fun.
• You have a public artifact — GitHub stars, a blog, X posts that engineers quote, a YouTube channel where you build in public.
• You contribute to open-source AI tooling, or to infra-as-code / observability projects.
• You can show evidence of doing the work of five people in a quarter — and explain how.
• You’ve helped a team replace a manual ops process with an agent that now runs on its own.

What the first call looks like
This is non-standard, so we’ll be transparent up front. The first technical call is one hour, hands-on, and runs entirely off your screen. Come ready to walk us through:
• Your AI productivity setup. The tools, the agents, the workflows, the hotkeys. We want to see your terminal, your editor, your prompt library, your custom skills.
• Your token usage history. Yes, actually open the dashboard on the call — we want to see what you actually use, how much, on what.
• Your three favourite projects. Demo each one, walk us through the architecture, and tell us what was hard.
• Your learnings. What worked, what didn’t, where AI broke down for you, what you do when a model is the wrong tool.
If the answer to any of this is “I don’t really track that” or “let me take this offline,” this isn’t the role.

What you’ll get
• Real ownership of the AI layer on top of infrastructure that processes millions of payments a day. Your work is load-bearing from week one. A team that wants to be automated out of the boring parts and will back the bets you make to get there.
• Freedom to use whatever models, agents, frameworks, and tools work best for the problem in front of you.
• Working with leaders building Razorpay’s next decade — not maintaining the last one.
• Competitive compensation, calibrated for an engineer who can do the work of a small team.