Indian Addresses are an E-commerce Nightmare, But Meesho's LLM is Fixing That

For decades, India’s addressing system has been functioning on an unwritten logic that humans understand instinctively but machines struggle to decode.

In some parts of a city, addresses may be precise, detailing house number and street, while in other parts, they could read more like a set of directions such as “behind the Hanuman temple”, “near the medical store”, “opposite the grocery shop,” etc. Add multiple spellings for the same locality, regional conventions, and the absence of formal street or house numbers, and the problem becomes particularly difficult for digital commerce and logistics companies.

According to research by Santanu Bhattacharya, a data scientist at MIT Media Lab, poor addresses cost India $10–14 billion annually, ~0.5% of the GDP.

In Western countries, standardised address formats enable relatively accurate geocoding, keeping last-mile delivery costs at around 10–12% of the total delivery cost. In India, however, last-mile costs can account for nearly 30%, despite the country’s lower labour costs, per the research. 

For e-commerce platform Meesho, whose customers primarily come from outside India’s largest cities, that challenge is harder to overcome. It has built its Geo India system specifically to understand the way Indians actually describe locations rather than forcing them to conform to globally standardised address formats.

Pratik Kumar, Head of Engineering - Monetisation, Payments, and Logistics at Meesho, explains in a conversation with AIM that “rather than teaching our users how to talk to machines we decided that we'll teach machines how to understand” Indian addresses.

India’s Non-Standard Addresses

Conventional geocoding systems typically work well when addresses follow a predictable hierarchy such as house number, street, locality, city, and postal code. Indian addresses are an outlier.

“Indian addresses are very different from when you compare it with all the global addresses,” Kumar explains. While addresses elsewhere generally follow some degree of standardisation, Indian addresses can be highly contextual, packed with landmarks, spelling variations, and locally understood references.

For someone familiar with the area, that may be perfectly adequate. For a machine, however, the description can be difficult to interpret.

That gap has direct consequences for logistics. Incorrect geocoding can lead to misroutes, failed deliveries, higher transportation costs, and ultimately a poor customer experience.

Kumar says that Geo India, combined with its Network Intelligence System, a machine learning-based framework to optimise supply chain operations, has already reduced wrong-route deliveries by more than 50% on heavy-shipment lanes. The company also says route-related costs have fallen by 5%, while address accuracy on these lanes has improved by 20 percentage points.

“Every successful delivery helps us in terms of making our model better,” Kumar says, describing the system as a self-learning loop. Each delivery generates can help the model understand an address, landmark, or location more accurately the next time it encounters a similar description.

Teaching AI to Understand 

Meesho designed the technology stack around the peculiarities of Indian addresses rather than adapting a conventional global geocoder.

Traditional geocoders often rely heavily on lookups based on standardised address hierarchies. Geo India instead uses a sequence-to-sequence approach to transform an Indian address written in natural language into a location that can be represented on a map.

The system also uses geospatial indexing, effectively dividing the earth's surface into a grid of hexagons and predicting which geographic cell an address belongs to.

But one of the more important components is its custom tokenizer.

Generic tokenizers can struggle with the spelling variations and landmark references common in Indian addresses. Meesho's custom tokenizer has been trained on 67 million Indian addresses, allowing it to recognise contextual references and different spellings for the same place.

For Meesho, that matters because 88% of its users are from rural India or outside the country's top eight cities, according to Kumar.

“Rural India is completely untapped,” he says, highlighting the gap between the way people naturally describe their locations and the formats expected by conventional digital addressing systems.

The company has now deployed the system across 100% of its deliveries.

AI is Also Mapping India

The significance of address intelligence goes beyond correcting existing addresses. India's physical landscape is changing rapidly, particularly around the edges of cities. New residential clusters, roads, and localities can emerge faster than conventional maps are updated.

“Typically, rural India is getting urbanised faster than any map that can keep up with it,” Kumar says.

He points to an example from the outskirts of Jaipur, where a new colony had sprouted, and people had started placing orders. Geo India detected the emerging cluster through delivery signals and passed that information to the Network Intelligence System for better route planning.

The long-term ambition is not simply to identify where a parcel needs to go. By combining Geo India with network intelligence, Meesho expects to improve estimated delivery times, capacity planning, and route selection.

“Geo India started as a map reader. But we see that as becoming a GPS for Indian commerce,” he affirms.

National Address Intelligence Layer

The broader opportunity extends beyond one e-commerce platform.

Sajid Malik, Chairman and Managing Director of Genesys International Corporation, states that India's fragmented addressing system needs a common digital foundation.

“India’s addressing ecosystem has historically evolved around local conventions, landmarks and regional formats, making it difficult to create a consistent digital representation of location at a national scale. The emergence of DIGIPIN can help change this by creating a common digital addressing layer and making address intelligence an important next layer of India’s Digital Public Infrastructure.”

But simply assigning every location a digital identifier may not be enough.

“India needs national standards for how locations are digitally represented, while allowing citizens to continue describing their addresses in the way they naturally do,” Malik says.

That is where AI could become the connective layer. Malik comments that AI and geospatial intelligence can interpret diverse and unstructured address descriptions and convert them into standardised location references.

“Once this common location layer is established, it can significantly strengthen geocoding, navigation, and last-mile connectivity while enabling richer location intelligence to be built around every coordinate,” he notes.

For e-commerce and retail companies, this has implications beyond minimising failed deliveries. Better location intelligence could expand serviceability and improve last-mile efficiency.

But AI adoption in logistics also needs to make economic sense.

Meesho found that relying heavily on external cloud AI providers could make some applications prohibitively expensive. That led the company to develop its own machine-learning stack, reducing its AI costs by 60–70% compared to cloud AI providers, Kumar notes.

The company is currently focused on solving Meesho's logistics challenges rather than opening Geo India as a general-purpose addressing service for Aadhaar, DigiLocker, India Post, or other platforms. But that could change.