Our Terms & Conditions | Our Privacy Policy
Rajeev Rastogi, Vice President, Machine Learning at AmazonMessy addresses, misspelt queries, and millions of first-time internet users with varying levels of digital comfort. India may be one of the most challenging markets to build AI for, especially in retail. And perhaps that’s exactly why Amazon is stress-testing some of its most ambitious AI systems here. From conversational shopping assistants and AI-generated listings to defect detection in mango crates, the company is embedding GenAI and machine learning (ML) deep across the customer, seller, and logistics journey.
Being a large country with multiple languages and dialects, queries often come in different phonetic forms such as “blutut sant systam” instead of ‘Bluetooth sound system’. Amazon’s stack tackles this using phonetic transliteration across eight Indian languages and semantic search to map noisy input to actual catalogue items.
“Our systems are trained to recognise phonetic variations and transliterations so that customers can type words the way they sound in their native language,” Rajeev Rastogi, Vice President, Machine Learning at Amazon tells ET Enterprise AI. This has resulted in higher search success rates and product discovery for non-English or transliterated queries, he adds.
However, with any model all the answers are not available from day one, it’s an “iterative process”. The first version of phonetic search, for instance, worked well on common transliterations but missed rarer regional variations. “By analysing customer interaction data and incorporating more diverse training inputs, we enhanced the model,” says Rastogi. This process reinforced the value of launching with a working solution, closely monitoring usage, and then rapidly improving based on actual customer behaviour, he adds.
Keeping it fresh
Fresh produce is another frontier where AI/ML is making a dent. At Amazon Fresh, crates of mangoes, tomatoes, and other fruits and vegetables are scanned by computer vision models and IoT-enabled cameras to ensure quality. “The first model detects and counts each visible item in the crate. The second model identifies specific defect classes such as cuts, cracks, or pressure damage by analysing annotated images of produce. Both models are trained on millions of labelled images to recognise a wide variety of defects with high accuracy,” Rastogi highlights.
For example, in a crate of mangoes, the system can detect the exact number of fruits and flag those with visible blemishes or cracks. Similarly, for tomatoes, it can identify pressure damage or surface cuts that might affect freshness. “Once the system flags a defect, sellers can take immediate action to remove or replace the affected items before they reach customers. This helps maintain consistent quality and strengthens trust between sellers and customers,” he shares.
The India jigsaw puzzle
Addresses add a different layer of complexity to the India story. Many are incomplete, with missing street names or landmarks. To tackle this, Amazon created an Address Deliverability Score (ADS). By comparing an input address with similar ones successfully delivered to in the past, ADS predicts how “locatable” an address is. “We have seen improvements in address quality, route accuracy, reduced travel times for delivery associates, and faster order processing in fulfilment centres. These gains help lower operational costs while improving delivery speed and reliability for customers,” says Rastogi.
Behind the scenes, AI touches every part of Amazon’s logistics chain in India. Fulfilment centres use models to optimise inventory placement, box sizes, and even associate walking routes. In last-mile delivery, AI plans routes factoring in traffic and density. “Large language models are being explored to clarify delivery instructions and better match delivery addresses to real-world entry points, which is especially valuable in multi-unit buildings or less-mapped areas,” Rastogi shares.
These optimisations, combined with ADS, are reducing travel times, improving processing speed, and lowering operational costs while keeping deliveries reliable.
A smarter storefront
One of the most visible rollouts at Amazon is the AI-generated review highlights. “It summarises thousands of product reviews into concise points, helping customers quickly assess whether a product meets their needs,” Rastogi notes. Linked back to the original reviews, these highlights reduce the time customers spend scanning pages and build confidence in purchase decisions.
Then there’s Amazon’s conversational shopping assistant called Rufus, which shifts search from keywords to context. Instead of typing ‘Bluetooth headphones’ customers can just ask: “What are the best wireless headphones for running under ₹2,000?” The assistant interprets intent, references the product catalogue and trusted sources, and guides them with comparisons and recommendations.
Rastogi highlights that “Over 10 million Indian customers have used the conversational GenAI shopping assistant, and we have seen encouraging signals such as higher engagement with products surfaced through Rufus compared to generic search results, along with customers spending less time navigating between pages before making a purchase decision.
“These trends suggest that context-aware recommendations are improving discovery and reducing friction in the shopping journey.”
Amazon also adapts the shopping journey based on customer proficiency. Signals such as number of search queries and navigation behaviour indicate whether someone is a first-time or an advanced user. Low-proficiency users see tutorials, simpler navigation, and language options, while experienced users get personalised ads, ‘Subscribe & Save’, and richer widgets. “This tailoring improves engagement for first-time and less experienced users, helping them complete their shopping journey more confidently.It also ensures that experienced users see features and offers that are most useful to them, which supports stronger product discovery and conversions.” says Rastogi.
From catalog chaos to clarity
On the seller side, GenAI is streamlining what is perhaps one of the most tedious tasks, that is creating product listings. “Sellers only need to upload an image and a few descriptive keywords, and generative AI models help to create richer and more effective product listings by generating engaging titles and descriptions,” Rastogi explains.
For example, from the input “running shoes, men, size 9, lightweight, breathable”, the system generates: “Men’s Lightweight Breathable Running Shoes – Size 9, Ideal for Daily Training and Long Runs.” This reduces time-to-list, ensures consistency, and boosts discoverability, especially during peak events like Diwali or Valentine’s Day, he notes.
The next frontier
So where is Amazon placing its AI bets over the next year? Rastogi points to multimodal models that can combine insights from text, images, video, and more.
The pipeline spans the everyday to the inventive: reading nutrition facts from product images, interpreting handwritten prescriptions, spotting damaged goods, auto-adjusting product visuals, generating marketing creatives, and even suggesting furniture that fits the style of your living room.
The use cases under development span the mundane to the inventive: reading nutrition facts from product images, interpreting handwritten prescriptions, spotting damaged goods, auto-adjusting product visuals, generating marketing creatives, and even suggesting furniture that complements what’s already in your living room.
Amazon’s India experiments underline that scaling AI is less about flashy models, and more about solving for the messy realities of the market.
- Published On Sep 2, 2025 at 08:54 AM IST
Join the community of 2M+ industry professionals.
Subscribe to Newsletter to get latest insights & analysis in your inbox.
Get updates on your preferred social platform
Follow us for the latest news, insider access to events and more.
Images are for reference only.Images and contents gathered automatic from google or 3rd party sources.All rights on the images and contents are with their legal original owners.
Aggregated From –
Comments are closed.