Inside Flipkart’s OneTech push as it gears up for an India IPO

In an interview with Mint, he explains how this unified stack is modernizing Flipkart’s architecture for an initial public offering (IPO) readiness, quick commerce expansion via Flipkart Minutes, trust features such as fraud detection, warehouse automation, and fintech growth as the Walmart-owned firm shifts domicile to India for a potential listing.

The Walmart-owned firm got the National Company Law Tribunal's nod this month to shift domicile from Singapore to India ahead of a potential 2026 listing. It follows Flipkart’s $1 billion funding round in May 2024, led by Google that valued it at $35-36 billion. Edited excerpts from the interview:

How is OneTech structured and how does it function? What are its core groups?

OneTech is the organization that delivers all products and services for every Flipkart consumption experience we ship to customers. It also supports all our internal business operations and corporate functions.

In terms of roles, on the product and platform side, we have a demand-side function that looks at all consumer experiences—everything about buying products and getting them delivered. We call this the consumer experiences team.

Then we have a supplier experience team that helps our sellers and suppliers do business with us. In between, a marketplace team handles cataloguing supplier inventory, understanding consumer demand, and matching demand to supply so that buying and fulfillment can be completed. We also have a supply chain organization that moves products through the network and gets them delivered. To wrap it all together, there is a fintech and payments organization that handles money movement between buyers and sellers.

Reports say Flipkart is increasing AI investments six-fold. Where is this capital going? Are you building your own models or leaning on OpenAI, Google and others?

AI is already deeply embedded across Flipkart’s experience ecosystem, delivery stack, marketplaces and internal tooling. Two or three years ago, we relied primarily on traditional machine learning models. With the advent of GenAI, we have upgraded many of these into what we call deep models, and we now combine large language models (LLM)-based technologies with SLM-based (small language model) technologies as well.

On the question of whether we use Gemini, ChatGPT, or our own small language models—it’s a combination of everything, depending on the use case. For example in our in-house LLM use cases, we utilize it to understand the user’s query or conversation (voice search). Once we grasp the context, we ‘fork’ tasks into specifically tuned small language models so they execute reliably, without hallucinating.

Specifically for search and discovery, this starts with understanding who you are including your past behaviour, purchase and browsing history, and intent, to personalize results and rankings.

When you say you’re “training models", are you training open-source models, or do you have fully in-house models based on Flipkart data?

We have hundreds of models; it depends entirely on the use case. It’s both: a mixture-of-experts setup and independent transformer-based models. We put to work whatever combination of intelligence is needed to meet a specific use-case outcome.

For rich media on the search results page, we rely on models for image generation, video stitching and related tasks. For catalogue quality, we use models for text schema understanding and data validation to ensure correctness. The model choice is always driven by the use case.

Tell us more about Flipkart’s voice search. What’s the actual adoption of voice for discovery? Is it mainly a tier-2 or tier-3 phenomenon or is also used in affluent metros?

Flipkart is going to make voice a first-class interaction mode. Historically, voice models weren’t mature enough—challenges included poor noise cancellation, inaccurate voice-to-text, weak language coverage, misinterpreted tones and slang. Now, with significant maturity in voice models and multimodal systems, voice plus text plus search will become central, especially for conversational commerce.

We expect users to literally have a conversation with the app, and to make the experience feel more like talking to a store assistant in an offline shop. Today, every text-based search or click is a separate session. In voice-based search conversations, we maintain the full context and memory from the start, so as the session progresses, responses get sharper and more personalized.

Flipkart serves shoppers across income levels. How do you segment your customer base, from metros to smaller towns, and tailor the app experience for each?

We operate strongly in the lower part of upper middle class and the lower middle class, and we’re pushing further down. At the same time, we must personalize both catalogue and experience for each segment.

Take Flipkart Minutes (quick commerce): it primarily targets metros with high density, where accessibility to stores is constrained by traffic and where people place a high premium on time and convenience. That pushes us towards the higher-affordability segment and higher ASPs (average selling price).

At the bottom, we have Shopsy, which competes with players like Meesho and targets the lower-income, value-conscious segment that will grow with us into the mid-market.

So at the top end, we compete with Blinkit and Amazon; at the very bottom via Shopsy; in the middle market, which is our core, we are effectively the leader. Because of that mid-market strength, we have roughly 48% of India’s e-commerce market share.

On quick commerce, Flipkart has spoken of adding around 1,000 dark stores. What are you solving for at a tech level to avoid bottlenecks as you scale?

There are multiple layers to this. First, we essentially remap India’s 22,000 pin codes into smaller polygons of varying sizes. At the smallest level, we look at each polygon and understand population density, historical demand and selection patterns (from our e-commerce business), external market research on demand, and local infrastructure conditions.

When these three—demand, population, infrastructure—hit an optimal model that meets consumer expectations, we can assess the need to open a good dark-store location. Second, for speed and resilience, we’ve mapped the country so that if one dark store can’t fulfil an order, we can intelligently check adjacent stores.

If demand forecasting is off, or one store runs out of a popular SKU (stock keeping unit), we can expand the service radius for that customer, including more dark stores to increase selection. That might mean slightly more delivery time, but it ensures availability, sometimes even when competitors have stocked out in that micro-region.