|
Listen to this story
|
In the past decade, fintechs have spearheaded a transformative shift in the way payments are conducted in India. These companies have played a pivotal role in driving the digital revolution across the country. One of the companies at the helm of it is Razorpay, a provider of payment gateway services to vendors, merchants, and e-commerce platforms.
Founded in 2014, the fintech unicorn has democratised access to modern payment technologies and, as a result, has empowered small and medium-sized merchants to participate in the digital economy and expand their operations.
Razorpay CEO Harshil Mathur once said that technology is the biggest differentiator for Razorpay. Today, in the age of generative AI, Razorpay is exploring multiple use cases of the technology. Analytics India Magazine recently caught up with Murali Brahmadesam, head of engineering & chief technology officer at Razorpay. Murali became a part of the fintech company in 2022, bringing with him over seven years of experience at AWS and over 10 years at Microsoft.
Murali notes that Razorpay was born in AWS in 2014. AWS excels in supporting startups by fostering a culture of experimentation and enabling the exploration of diverse ideas without immediate scalability concerns. This approach has, in fact, proven highly beneficial for Razorpay.
Exploring Generative AI
Razorpay’s exploration of generative AI capabilities is divided into two parts, according to Murali. The Bengaluru-headquartered firm recently held a hackathon on generative AI and the team came up with 12 different use cases for generative AI, he said.
“The second part, take, for instance, the extensive documentation that developers often have to navigate through in order to understand how to use a service or onboard onto a platform. This is where the conversational search capability of tools like ChatGPT can prove invaluable,” he said.
Further, the capability also extends to improving customer interactions across various channels, such as supporting tickets, chat, or email. “By summarising multiple conversations, agents can respond more effectively and provide a seamless and efficient customer experience.”
On Large Language Model
Generative AI became huge with the launch of ChatGPT, which is powered by GPT large language models (LLMs). Today, we have a host of LLMs developed by different AI Labs like Cohere AI, Anthropic and Stability AI, and Google, among others.
Large language models are trained on massive amounts of text data mostly scrapped from the internet. However, an enterprise does not need its chatbot to have all the knowledge of the world. Murali believes domain-specific language models trained on specific datasets might be more useful for enterprises. Not everyone can build an LLM, according to him.
“For Razorpay, the question is should we use an externally hosted service or build domain-specific LLMs that can help us. We are exploring all possibilities and I think we will have a mix of both. We are enthusiastic about the emergence of open-source frameworks such as LLaMA, Dolly 2.0 etc. We are actively exploring these frameworks to assess their potential benefits and functionalities,” he said.
Razorpay is carefully evaluating each case individually, considering factors such as the cost of rebuilding, hosting requirements, and overall feasibility. “This allows us to make informed decisions about which parts should be rebuilt, self-hosted, or integrated as part of our service offerings.”
How has Razorpay been using AI?
Even though Razorpay is exploring generative AI use cases, the fintech firm has already been using AI for different purposes.
“One of the key use cases involves using AI to determine the most suitable bank when processing payment requests. With hundreds of options available, AI helps make informed decisions quickly and efficiently. This is particularly crucial in time-sensitive scenarios where low latency is essential,” he said.
Razorpay also uses AI to understand consumer behaviour. The AI takes into account various factors such as the likelihood of return and this helps the merchant make a call on whether they should provide ‘Cash on Delivery’ as an option in that particular case.
Furthermore, the company also uses it for fraud detection. Called Third Watch, now integrated with RazorPay Magic Checkout, the tool helps merchants to profile fraudulent transactions and detect them in real-time. The algorithm analyses over 300 parameters, such as shopping behaviour, location data, and other transaction features, and identifies patterns that indicate potential fraud, he concluded.
