How Simpl is leveraging AI & ML to enhance fraud detection
With more than 20,000 merchant partners and 25 million users across India, Simpl is focused on ensuring seamless and safe transactions between users and merchants. Sheekha Verma, Director, Data Science, Simpl shares about the company's robust anti-fraud infrastructure, including building data science systems
What are the key features of your robust anti-fraud infrastructure?
Anti-fraud infrastructure is not a model or just one monolith system. It comprises multiple layers, from simple heuristics filters blocking / flagging transactions to ML models, deep learning models, and graph network models that identify anomalies in transactions and user behavior. We have a robust team of engineers and analysts dedicated to anti-fraud infrastructure. Additionally, we have a dedicated team of fraud monitoring colleagues who look at all high-risk, flagged transactions and, if needed, call them to verify account/transactions.
How are you using AI & ML to enhance fraud detection?
Simpl, in its six years, has seen 150+ million unique users on its 20,000+ merchant partners to date. We leverage deep learning models like LSTM to identify unusual changes in a user’s shopping behaviour. We also have bipartite graph networks scoring evolving clusters and connections between users. This helps us identify bot networks, simulators, or even groups active on specific merchants like taxi rides, grocery, digital goods, etc.
Please explain how your systems are ensuring seamless and safe transactions between users and merchants.
Our automated systems validate that the transaction associated with a user is authenticated. We look out for sudden changes in location, amount patterns, etc. If needed, our operations team call to validate if the user did the transaction.
We’re also building systems to ensure the merchant is valid in cases of aggregator platforms. We continuously monitor spikes in transactions from a location, on a merchant, etc., to identify new patterns and fraud attempts.
Fraud is forever changing. As we build systems to tackle one kind of fraud, we always observe a shift in pattern after some time. Our intent is to have a complete system with multiple levels to authenticate and block accounts.
What are the common fraud scenarios in BNPL space, and which best practices have you adopted to counter these challenges?
There are multiple kinds of fraud. Account takeover, identity theft, and chargeback frauds– most common in the fintech industry, and the way to identify those are now almost standardised. What is interesting is the case of organised fraud or cluster fraud. This has increased post-pandemic. In fact, a report by CIBIL also shared how organised fraud is now moving to e-commerce.
Additionally, in India, the case of sharing OTP unknowingly to get discounts, etc, is also prevalent. There are social media groups where OTPs are traded. For us, we treat them similar to cluster and community fraud.
What are your upcoming plans for scaling up fraud detection capabilities and infrastructure at Simpl?
We have grown more than 3X in the last six months. That means our infrastructure handles a much higher load while our SLA requirements to the front-end team remain < 100 ms. To manage the concurrent peak RPM and the latencies of <100 ms, as we also build multiple systems to review transactions in real-time, we will be taking up a re-architecture project in the next quarter. Additionally, we will also be experimenting with technologies like tiny ML.