Razorpay 2026 AI Engineer Roles: What Freshers Need to Know

Razorpay’s engineering team ships production ML at payments scale: fraud detection models, risk-scoring engines, and multi-agent systems that decide in milliseconds whether a transaction clears.

Most of this infrastructure is less than four years old. The company is still building it, which means the engineers who land SDE-1 and ML roles in 2026 will work on systems that don’t have a finished shape. That’s a different kind of role from joining a team that’s in maintenance mode.

What Razorpay Actually Builds

Razorpay is a payments gateway and financial infrastructure platform serving businesses across India. The core product handles the moment of transaction, but the technical surface extends to credit infrastructure, payroll, banking APIs, and merchant analytics.

Every one of those surfaces touches a risk question: is this transaction legitimate? Is this merchant operating within policy? Is this authentication pattern consistent with prior behaviour?

The ML organization answers those questions at production scale. Payment decisions happen in well under a second. Fraud patterns shift daily. Authentication success rates connect directly to merchant revenue, because a failed payment costs the merchant the sale. That’s why Razorpay’s engineering blog is dense with ML systems content and why the ML team is one of the company’s primary growth areas going into 2026.

The scale of the problem is what makes Razorpay interesting from an engineering standpoint. You’re not building an ML demo. You’re shipping models to production, measuring their impact on transaction success rates, and iterating against a live payments network.

Where AI Shows Up in Razorpay’s Products

Four areas show up clearly in Razorpay’s 2025 engineering output.

Merchant Fraud Detection: Bumblebee

Bumblebee is Razorpay’s flagship agentic AI system. Published as a technical deep-dive on DEV Community in 2025, it evaluates merchant websites for fraud risk. Before Bumblebee, this review was largely manual. The system now processes each merchant evaluation in under 90 seconds, replacing approximately 800 hours of monthly manual review work. Detection accuracy improved from 88% to over 99%.

Bumblebee is worth studying beyond its headline numbers. The architecture uses an orchestrator that delegates to specialized sub-agents, each handling a different dimension of merchant risk. That pattern, an orchestrator coordinating specialized agents, is the same one appearing across enterprise fintech AI deployments. Understanding why it works (and what its failure modes are) is more useful interview preparation than memorizing accuracy statistics.

Payment Authentication: The ACS Risk Engine

Razorpay’s ACS (Access Control Server) risk engine achieves up to 95% authentication success while keeping fraud losses within acceptable limits. This is a production ML system, not a prototype. The engineering blog posts behind it cover feature engineering on transaction signals, model calibration, and the deployment constraints typical of high-stakes financial ML. A candidate who reads them before an interview will have a concrete vocabulary for the problem domain.

Agent Studio

In 2025, Razorpay launched Agent Studio, a platform that lets businesses build custom AI agents on top of Razorpay’s payment and risk infrastructure. The internal agentic capabilities Razorpay built for its own operations became a product layer for merchants. For ML engineers inside the company, it means the infrastructure built internally eventually ships as a feature. That’s a different incentive structure than pure internal tooling.

Payment Routing Optimisation

ML models at Razorpay predict the optimal payment route across banks and processors in real time. Success rate optimisation is among the highest-stakes ML problems in fintech because a failed payment has an immediate, measurable cost. The model doesn’t need to be perfect; it needs to consistently outperform the prior routing heuristic.

The Three Fresher Entry Paths

Track	Who Is Eligible	Selection Stages	CTC (Aggregator Estimate)
SDE-1	CSE, IT, ECE freshers — campus and off-campus; 60% aggregate required	Online coding test, 2 technical rounds (DSA + LLD), hiring manager round	14 to 20 LPA
ML Engineer	Prefers 2–3 years; IIT/NIT freshers with NLP or agentic AI research sometimes considered	ML portfolio screen, ML system design round, Python coding round, leadership round	25 to 45 LPA
Software Engineer Intern	Penultimate and final-year students; CSE, IT, ECE	Online coding test, 1 DSA interview, culture fit conversation	60,000 to 80,000 INR per month

CTC note: Razorpay does not officially disclose compensation figures. All values above are aggregator estimates and should be treated as directional, not authoritative. Actual offers vary by cohort, branch, and negotiation.

For most freshers, the SDE-1 track is the realistic starting point. The ML Engineer path is largely closed without research output: a published paper, a deployed dataset, or a production system (even a personal project with real API calls) changes the profile. The intern-to-FTE route offers a lower barrier of entry: one technical interview round instead of two, followed by performance over the internship duration.

Razorpay accepts both campus and off-campus applications via razorpay.com/jobs. The campus pipeline is strongest for Bangalore-region colleges and IITs/NITs. Off-campus applications follow the same selection process.

What the Selection Process Actually Looks Like

Stage 1: Online Coding Test

The test focuses on DSA: arrays, graphs, dynamic programming, and tree traversal are the recurring topics. Python and Java are the accepted languages. Difficulty is comparable to medium-level problems on standard coding platforms. Time management matters more than language choice; picking the one you reason fastest in is the right call.

Stage 2: Technical Interview Rounds

SDE-1 candidates face two rounds. The first covers DSA, similar to the coding test but with the interviewer expecting a running explanation of decisions at each step. The second covers low-level design (LLD): class structures, object relationships, and basic system design. Razorpay’s own products provide natural LLD prompts: payment state machines, merchant onboarding workflows, event-driven transaction systems. Familiarity with the domain helps construct plausible designs under time pressure.

ML Engineer candidates skip to an ML system design round after the portfolio review. Expect questions about feature stores, model serving architecture, and drift detection. The Python coding round tests ML algorithm implementation, not just theoretical knowledge.

Stage 3: Hiring Manager Round

This is a structured conversation about project experience, problem-solving approach, and fit with Razorpay’s engineering culture. For freshers, it typically centres on how they worked through a difficult technical problem during a project or internship. Being able to walk through a system you built end to end, including the decisions you’d make differently now, is worth more than a flawless recitation of theory.

Building the Profile Razorpay’s ML Teams Look For

The engineering blog at engineering.razorpay.com is genuinely useful interview preparation. The posts describe the actual technical decisions behind Bumblebee, the ACS engine, and the distributed systems running payment routing. Knowing what problems Razorpay has already solved, and how, gives interview answers a domain grounding that generic ML prep does not.

For the SDE-1 track, the skill overlap with standard placement prep is high: DSA, LLD patterns, and a working understanding of distributed systems basics. The AI-specific angle matters more for ML roles. Even SDE-1 candidates, though, benefit from understanding how Razorpay’s ML surfaces work, because the hiring manager round often goes in that direction.

The AI hiring context at Razorpay differs from what enterprise IT companies are doing. Accenture’s 2026 AI hiring shift, for instance, adds AI screening layers onto established fresher tracks. At Razorpay, the ML work is embedded in the core product, not layered on top, which changes both what the interviews test and what the day-to-day engineering work looks like.

In November 2025, Razorpay hired Prabu Rambadran, formerly at Google Cloud, as VP Engineering to lead AI and engineering expansion. The senior hire signals the direction: Razorpay is building toward an AI-native engineering organisation, and the roles opening up in 2026 reflect that trajectory.

The broader skill sequence from placement-ready fundamentals to the ML engineering layer that companies like Razorpay are hiring for is mapped out in the 2026 AI roadmap for Indian engineering students. It’s the right companion read if you’re building toward a fintech ML role and need to sequence your preparation.

Razorpay’s ML team tends to interview by asking candidates to walk through an agentic system they’ve built. A deployed project with real API calls is a more credible signal than any certificate. TinkerLLM is where you build that first project: at ₹299, you get working LLM API calls and a deployable micro-project you can walk through in an ML interview without hand-waving about what the system actually does.