How AI-powered address intelligence can shape the next chapter of logistics in India
Every e-commerce journey begins with understanding an address. Even before an order is placed, location intelligence helps determine how it can be served and sets the journey in motion, guiding the parcel through the logistics network to the customer’s doorstep.
India’s geography is extraordinarily diverse, and so is the way people describe it. An address can be a house number and street name, or it can be a description built around landmarks, local references, neighbourhood names and directions that are familiar to someone living in the area. As e-commerce expands across Bharat, understanding these descriptions with greater precision is becoming an increasingly important part of building logistics networks that can serve India at scale.
A customer in a major city may enter a structured apartment address, while a customer in a smaller town or village may describe their location as “the house after the water tank, beside the government school, near the main road.” The same address can be written in different ways, across different languages and with variations in spelling and local jargons. For technology to work effectively across this, it needs to understand the nuances embedded within these descriptions and connect them to the physical geography they represent.
This is where the next frontier of logistics intelligence lies: AI that can understand the nuances of India’s locations and translate them into precise, actionable geographic intelligence.
The unique complexity of India’s address system
For a logistics network, an address has to be translated from human language into a precise geographic location or set of coordinates in the physical world. In India, doing so requires an understanding of the context contained within the address itself.
A landmark may only carry meaning within a particular locality. Regional vocabulary can add layers of geographic context, while two very different descriptions may ultimately point to the same destination. The challenge lies in recognising these relationships and understanding how language connects with geography.
This requires machines to learn from the patterns that emerge across addresses and locations. They need to recognise local references, understand spatial relationships and identify how different expressions can represent the same physical place.
As logistics networks expand across Bharat, this makes precision along with coverage increasingly important. Reaching a PIN code is only the beginning. Within it are thousands of individual destinations, each with its own geographic and contextual markers.
From geographic coverage to geographic intelligence
This is where AI can add a new layer of intelligence to logistics. For decades, much of logistics technology has focused on optimising what happens once a destination is known. In a country spanning more than 19,000 PIN codes, the challenge extends beyond knowing which region or PIN code a shipment needs to reach. It requires understanding the specific location within it. AI is now creating an opportunity to extend intelligence further upstream, helping systems understand the address itself before the rest of the logistics journey begins.
Solving this requires AI to learn two languages simultaneously: the language people use to describe a place, and the geography those words represent.
When trained on large volumes of real addresses paired with verified locations, AI systems can begin to recognise the relationships that make an address meaningful. A landmark becomes a geographic signal. A variation in spelling becomes part of a recognisable pattern. A local reference can be understood in relation to the geography surrounding it.
At scale, this capability can become increasingly powerful because a logistics network itself generates a continuous stream of location intelligence. Every successful delivery, verified destination and corrected address can become a learning signal, helping systems build a richer understanding of how people describe locations and how those descriptions map to the physical world.
Over time, this creates a continuously evolving layer of intelligence that can help logistics networks understand and serve India’s diverse geographies with greater precision.
Teaching AI to Understand How India Gives Directions
This was the challenge Meesho set out to solve with GeoIndia LLM, a purpose-built address intelligence system designed around the way locations are described across India. We framed this challenge simply: rather than teaching our users how to talk to machines, we decided to teach machines how to understand Bharat.
This system was trained on millions of Indian addresses, drawing on millions of real delivery traces across thousands of PIN codes, where customer-written addresses were paired with verified delivery locations. This enabled the model to learn the relationship between language, local context and geography, including landmarks, neighbourhood references, regional terminology and variations in how the same location may be written.
At its core, GeoIndia LLM combines language modelling with geospatial intelligence. Imagine the country divided into millions of smaller geographic polygons. The system uses different layers of intelligence to first identify the right geographic area and then narrow down the precise location within it.
The concept is somewhat similar to how PIN codes help organise geography, but with far greater precision. A single PIN code, particularly in rural India, can span kilometres and cover multiple towns, villages and localities. Identifying the right PIN code, therefore, does not always provide enough information to accurately locate a customer. GeoIndia LLM goes several layers deeper, translating an unstructured address into increasingly precise geographic cells and using the context within the address to identify where the customer is actually located.
This allows the system to interpret landmarks, neighbourhood references, regional terminology and other spatial cues, connecting the way people describe a location with the physical geography it represents.
The model was designed to account for the diversity of India’s address patterns, with region-specific models capturing local characteristics across different states. It also continues to learn from successful deliveries, corrected addresses and verified destinations, creating a feedback loop where real-world delivery outcomes strengthen its understanding of neighbourhoods, landmarks and regional address patterns over time.
Deploying this intelligence at scale required significant engineering optimisation. Meesho reduced GeoIndia LLM’s latency from approximately 700 milliseconds to around 80 milliseconds, enabling the system to operate within live logistics workflows at scale.
From Understanding an Address to Moving Millions of Shipments
The value of precise location intelligence extends far beyond identifying a destination. At Meesho, GeoIndia LLM feeds into the Network Intelligence System, which uses location intelligence alongside demand forecasts, shipment volumes, network topology, capacity, delivery timelines, service reliability and cost, to optimise how shipments move across the network.
This connects two fundamental layers of logistics intelligence. Understanding where a shipment needs to go creates the foundation for determining how it can move through the network efficiently.
The impact is already visible across Meesho’s logistics operations. GeoIndia LLM improved geocoding accuracy by 20 percentage points across shipping-heavy lanes and reduced misroute-related costs by 5%. Working alongside the Network Intelligence System, it has also contributed to reducing last-mile misroutes by more than 50%.
Benchmarked against a leading mapping system, GeoIndia LLM also reduced average location error by more than 50% and improved accuracy for difficult addresses by more than 85% across multiple states, including locations the system had not encountered during training.
As commerce continues to reach more towns, villages and localities across Bharat, the intelligence required to serve these geographies will become increasingly sophisticated. GeoIndia LLM represents Meesho’s approach to this opportunity: building AI that can understand the local context behind an address and translate it into intelligence that can power decisions across the logistics network.
For Meesho, GeoIndia LLM is an important step towards building that capability. By learning from millions of real delivery journeys and translating the way people describe their locations into actionable geographic intelligence, it is helping build a logistics network designed around the scale, diversity and realities of Bharat.
After all, every parcel may travel through a sophisticated network, but every journey still ends at a single address. As AI learns to understand India’s addresses with greater depth and context, location itself can become a powerful layer of intelligence powering the journey from one doorstep to another.
Authored by Pratik Kumar, Head of Engineering, Meesho