Inside Meesho’s big AI bet: Why India’s most-downloaded shopping app is betting on infrastructure

When India’s e-commerce companies talk about AI, the conversation usually gravitates toward chatbots, recommendation widgets, or a shiny new “assistant” bolted onto the app. Meesho’s FY26 annual report — its first as a listed company after debuting on the NSE and BSE in December 2025 — tells a different story. Buried inside letters from its founder – CEO Vidit Aatrey and CTO Sanjeev Kumar is a thesis that reads less like a marketing pitch and more like an infrastructure roadmap: build the foundational AI stack first, and let the features follow.

That bet has a name — BharatMLStack — and it sits at the centre of almost everything Meesho says about technology this year.

The Foundation: BharatMLStack
BharatMLStack is Meesho’s in-house machine learning platform, and the company describes it as the layer that “powers intelligence across discovery, advertising, fraud prevention, cataloguing, and customer support.” The pitch for building it in-house rather than leaning on third-party cloud AI services is squarely about cost and speed at scale: Meesho says the stack runs at 60–70% lower inference and AI workload costs than equivalent cloud services, while standardising how machine learning systems are developed and deployed so different teams aren’t reinventing the wheel.

It’s the same logic Indian banks have been applying to their own in-house GenAI platforms lately — give every team shared, governed infrastructure instead of forcing each department to build (and pay for) its own AI stack. For a company whose entire value proposition rests on razor-thin margins and “Everyday Low Prices,” controlling the unit economics of AI inference isn’t optional — it’s existential.

As CTO Sanjeev Kumar puts it in his letter to shareholders: “the biggest barrier to participation in commerce is expertise and not access to the internet.” He frames the shift plainly: “For much of history, e-commerce rewarded expertise; now software is making it increasingly optional, shifting from giving tools to doing tasks for them.”

Solving for Addresses That Don’t Exist on a Map
The clearest illustration of Meesho’s “build for India, not around it” philosophy is GeoIndia LLM, a proprietary model trained on millions of real delivery traces across thousands of pin codes. The problem it solves is deceptively unglamorous: in much of India, addresses aren’t street names and house numbers — they’re directions relative to a temple, a school, or the third lane past the tea stall. Conventional geocoding systems choke on that kind of input. GeoIndia doesn’t; it converts vernacular, landmark-based addresses into precise delivery coordinates, and every corrected address feeds back into the model to compound its accuracy over time.

The results, per the report: a 20-percentage-point improvement in geocoding accuracy and a 5% cut in misroute-related operational costs over the last year. The work was significant enough to be published at CIKM 2025, where Meesho says it outperformed industry-standard commercial geocoding systems across every evaluation metric it was tested against.

Kumar describes the underlying design principle behind that model, and Meesho’s engineering approach more broadly, as: “start with the reality of the user and build technology around it,” adding that “technology should adapt to how people actually live rather than expecting people to adapt to technology.”

On the logistics side, a companion system called the Network Intelligence System (NIS) handles ML-powered route planning — optimising truck allocation, sequencing, and delivery timing together, and continuously relearning as network conditions shift.

Making Commerce Conversational
Perhaps the most telling admission in Kumar’s letter is that Meesho’s biggest barrier to growth was never internet access — it was expertise. Search bars, filters, and category trees assume a user already knows how to shop online. Offline shopping in Bharat doesn’t work that way; it’s conversational. As he recalls:

“We often watched users hesitate after opening the app — they knew what they wanted, but translating that into search queries and filters didn’t come naturally.”

That belief produced Vaani, Meesho’s voice AI shopping agent, which launched in Q4 FY26 and crossed 1.5 million users in its first month alone. Users who adopted it showed a 22% lift in conversion, and — notably — first-time shoppers who had previously abandoned the app mid-purchase were able to complete transactions through voice. A sibling product, Chorus, extends the same idea to sellers: a voice-AI agentic platform that guides sellers through promotional events and now handles up to 300,000 calls a day autonomously.

On the discovery side, PRISM (Personalised Ranking and Intent Signal Module) blends long-term preference data with real-time behavioural signals; the report notes that over 75% of orders on Meesho now originate from personalised feeds rather than active search. A companion feature, Trendpulse, is described as an LLM-powered engine for surfacing regional and emerging cultural trends before users search for them.

Trust at a Scale Humans Can’t Review
With 166 million active listings and billions of product impressions, manual moderation was never going to be viable. TrustMesh, Meesho’s deep-learning integrity model, reasons across user behaviour, shared identities, and network relationships to catch fraud and abuse, and to predict return-to-origin (RTO) risk before a package is even dispatched. In FY26, the report says TrustMesh blocked roughly 9 million high-risk transactions and restricted about 2 million consumers and 62,000 sellers, cutting RTO by more than 10% since deployment. Like GeoIndia, this work was peer-reviewed — published at AAAI 2026.

AI Writing Meesho’s Own Code
One line in the Management Discussion & Analysis section is easy to miss but says a lot about where the engineering culture is headed: over 70% of Meesho’s code is now AI-generated, which the company credits with letting it ship products faster and more reliably than at any point in its history. Combined with the Board’s Report noting stepped-up investment in large language models, agentic platforms, and “strengthening of AI/ML and engineering talent across the organization,” it’s clear this isn’t a side project — it’s being treated as core infrastructure spend, on par with logistics and cloud capacity.

Meesho has also formalised a dedicated research arm, Meesho AI Labs, tasked with building India-specific AI models for personalisation and product discovery and pushing agentic AI further into the post-order journey — the part of the shopping experience that today revolves around returns, refunds, and customer support.

The Numbers Behind the Bet
The scale this infrastructure needs to support is considerable. Meesho’s Annual Transacting Users grew 33% year-on-year to 264 million in FY26, up from 199 million and 156 million in the two years prior. Placed orders rose 45% YoY to 2,668 million, and Net Merchandise Value climbed 39% to ₹41,560 crore. That growth came at a cost, though — contribution margin as a percentage of NMV compressed from 5.6% in FY24 to 3.5% in FY26, and Adjusted EBITDA Marketplace margin widened its loss to -2.8% of NMV, which the company attributes partly to higher technology and AI infrastructure investment, including training multiple deep learning models and scaling LLM and agentic capabilities.

Infrastructure Over Features, Again
Strip away the product names and Meesho’s narrative rhymes with what larger, better-capitalised incumbents like HDFC Bank have been saying about their own AI strategy: don’t start with the flashy customer-facing feature, start with the shared, governed, cost-efficient infrastructure underneath it, then let discovery, fraud detection, logistics, and customer service all draw from the same well. HDFC calls its version Neev; Meesho calls its BharatMLStack. Both are explicitly framed as the “railroad” that has to exist before AI can meaningfully scale across an organisation.

The difference is what each company is optimising the infrastructure for. HDFC’s stack is built to embed AI into an existing 120-million-customer banking relationship. Meesho’s is built to solve a more foundational problem — getting several hundred million Indians who have never transacted online, in a language and format they’re comfortable with, past the barrier of participation altogether. Whether that bet pays off in margins remains an open question the FY26 numbers don’t yet answer. But Kumar’s own closing line in his letter leaves little doubt about the destination Meesho is engineering toward: “we are building towards a future where technology fades into the background and participation becomes the default, where consumers do not need to learn e-commerce and sellers do not need specialised expertise to succeed.” Not new features, in other words — new foundations.

-This article draws on data and disclosures from Meesho Limited’s Annual Report 2025–26, including the CEO’s and CTO’s letters, the Management Discussion & Analysis section, and the Board’s Report.