How Flipkart tech mapped India so well
Delivery executives of e-commerce companies don’t need to call you to identify your location anymore.
It’s a dramatic change for a country where addresses are so vague — near this temple, off that main road. This story is about how Flipkart solved the great Indian address conundrum, to deliver its more than 1-million shipments across all of the country’s 22,000 pincodes every day.
Mayur Datar, principal data scientist at Flipkart, says it took India more than 50 years to map just 2,575 kilometres of its landmass, a meagre 0.7% of the country’s total 3.2 million sqkm area. “However, technological developments have helped map the entire country within a decade from the early 2000s,” he says.
In other countries, where addresses are structured, mapping for deliveries is based on the latitude and longitude of a location. In India, it needed a lot more than just off-the-shelf satellite imaging. It needed high-end computer processing, natural language processing, machine learning and artificial intelligence.
Natural language processing and AI/ML algorithms are used to understand the address texts that users provide. Ravindra Babu, who was part of Flipkart’s data science team for five years and who recently moved to Flipkart Group company Myntra to head its data science team, says customers often provide elaborate addresses, running up to 200 words, in order to pinpoint their location. Often, different family members write their addresses differently, some may mention nearest landmarks.
But this process by itself does not provide the needed accuracy for many addresses. So Flipkart used delivery executives to mark out sub-regions with a name — such as Ram temple or banyan tree — and to label addresses around it for a better understanding of the region. The machine learning algorithm then assigns addresses to these sub-regions. And when an unseen address comes from a sub-region, the system is able to classify it.
The system is so good now that if someone tries to game Flipkart’s rules, it’s able to stop it. For instance, in certain high-value products, you may not be allowed to buy more than one item. But users try to overcome that by using different IDs and providing variations of the same address. “The system recogises that the addresses are same,” Babu says.