Tech infra is costly, frugality is a big part of strategy: Flipkart's Datar

Extracting efficiency and managing costs, while passing on the savings to customers, is what e-commerce at scale is about, says Mayur Datar, chief data scientist, Flipkart

Flipkart is one of the earliest players to have bet on artificial intelligence, which has permeated all aspects of life. The company’s chief data scientist, Mayur Datar, has been driving all AI-related initiatives at the firm since he joined it in 2015. In an interview with Aryaman Gupta and Shivani Shinde, he describes how the homegrown e-commerce giant is leveraging AI, machine learning and data science to offer India-specific use cases. Edited excerpts:

How has Flipkart’s AI/machine learning landscape evolved over the years?

When I say AI, it also means ML, analytics, statistics, and more. When I joined Flipkart in 2015, we had a small team, which has grown a lot since. AI can really help you to bring in efficiency and scale products. Especially when you are scaling up a product to millions of customers and sellers, it can become a costly endeavour. Extracting efficiency and managing costs, while passing on the savings to customers, is what e-commerce at scale is about. From day one, we have invested a lot in terms of manpower and hardware resources into data.

What are some specific AI use cases that Flipkart has incorporated onto its platform?

There are many. Image moderation is a good example. As a marketplace, we witness a lot of sellers upload images of their products to our platform each day. These images must be moderated and vetted for quality, offensive content and other compliances. These checks were previously done manually, when the team was small, and now we use computer vision to do this.

Our supply chain is also a key focus area. Accurate demand forecasting at various levels, in terms of individual products, categories or across geographies, has multiple vital downstream applications, such as seller feedback and inventory management. Figuring out optimal routes for last-mile deliveries is also an area where we use AI/ML.

Discovery is a very critical component of any e-commerce platform. To enhance product discovery, we have built algorithms specifically for the Indian market, which account for specific spelling mistakes and translation errors that are often made when dealing with Indian customers. Language detection, especially when translating vernacular languages into English, has also been a focus area. This has improved our search options and recommendations, including areas like auto suggestions and better ROI to sellers from advertisements.

Fraud detection and prevention is another huge use case. Cases of fraud are especially high in the Indian market. For instance, we have had cases of customers ordering, then returning iPhones in their original packaging after switching out the device. This can happen at multiple touch points. To mitigate this, we use X-ray-based fraud detection to check returned products, without opening the original packaging.

Getting proper returns on investment from AI/ML has been a challenge for many enterprises. How has Flipkart fared?

We have seen tremendous ROI from our bets in data science. This is largely due to maintaining a healthy portfolio of projects that we have worked on, focusing on both here-and-now projects that have brought in efficiencies, as well as future technology. This diversified portfolio has given us tremendous ROI in terms of our data science investments.

Is Flipkart incorporating Generative AI applications into its technology stack? If so, how?

We already have a chatbot for post-order customer services. Our aim there was to minimise transfer to human agents, as many customers tend to ask similar queries. We have also launched a product-specific bot which answers customer queries relating to particular products.

We have launched assistance bots, and are now working on ChatGPT-like offerings in the realm of content generation as well. This includes applications such as product summaries, generating higher-quality product images, creatives for merchandising and advertisements, and augmented reality-driven product visualisations, among others.

What are the challenges in building an effective AI model in India?

The first challenge is identifying the right opportunities and avenues for cost management and reducing loss-making areas. For instance, product returns in India are common, and so are cases of fraud. Accounting for such challenges is vital.

Secondly, India is an extremely dynamic market. The landscape of internet users in India has changed rapidly, along with the e-commerce market. We are not seeing steady growth, but exponential growth. Building for such an ever-changing system is a challenge. For instance, demand forecasting is contingent upon observing steady patterns. Finding ways to account for dynamic demand when there is scarcity of data is a big challenge in the Indian context.

When we started building our translation model, our content was predominantly in English, and then we decided to start building an in-house transcription platform. We noticed that for the Indian context, and Indian e-commerce, these translation products do not work well. Then we started our journey on creating translation models. Our translation models are far better than what is available. Lots of translation products are built on parallel corpora, which means the same sentence will be said in two different languages. But this is pathetic when it comes to Indian languages.

Talent acquisition has also been a hurdle. Onboarding data scientists and experts in AI/ML is difficult in India. Furthermore, maintaining proper margins is another impediment. In the Indian market, where margins are generally lower, technology infrastructure is expensive. We have much less hardware at our disposal. Frugality is, therefore, a big part of our strategy.

What is Flipkart betting big on, going forward?

Our mission remains the same. We are looking to find the best ways to incorporate emerging technology onto our platform, onboard new customers and make them return customers. We see a lot of AI/ML projects that we are currently working on, such as gen AI, carving out our path ahead. Technology is constantly changing and our focus is on adapting and staying agile.

Does Flipkart see any potential in drone deliveries?

We do see potential for drone deliveries in India. We have experimented with mid-mile drone deliveries in the past. But there are also many accompanying challenges, such as city planning and acquiring licences. However, we do see potential in this space.

Has AI helped Flipkart onboard new customers from Tier-II regions and beyond?

Over the last three to four years, we have been looking at ways to take e-commerce to Tier-II cities and beyond. A lot of new customers are coming to the Flipkart platform, thanks to the mobile revolution in India. These customers are different from their metropolitan counterparts, a lot of whom are mobile-first users and speak vernacular languages. Their trust in the platform is generally lower in terms of payments and products.

Making the Flipkart app available in these vernacular languages was our first aim. Today, our app is available in 11 languages aside from English. A bulk of the content on the app is translated, be it product descriptions, user reviews, or key value attributes.

Using voice assistance and video search interfaces further aids in product discovery for Tier-II+ customers. We have found that the uptake for our video search feature is higher in Tier-II regions and beyond, and even more so among women customers and in categories like lifestyle. Using our video search tools, customers can upload a picture of the product they are searching for and get similar results.

We were one of the earliest adopters of voice search. We began by offering voice assistance for grocery purchases and have branched out since. We are now experimenting with other interfaces where voice can be used.