Meesho open-sources in-house ML platform. What does it mean for Indian startups?

In a move aimed at democratising access to advanced AI infrastructure, Meesho has open-sourced key components of its internal machine learning (ML) platform, BharatMLStack, on GitHub. 

The release, which includes a production-grade feature store, control plane, orchestration UI, and SDKs, positions the IPO-bound company among the first major Indian e-commerce players to contribute at this scale to the country’s growing AI ecosystem.

Developed over the past two to three years, BharatMLStack was purpose-built to support a wide range of real-time machine learning applications across Meesho’s user and seller-facing systems. In fiscal year 2025, the platform processed an average of 1.91 petabytes of data per day, executed 66.9 trillion feature retrievals, and enabled 3.12 trillion real-time inferences during peak loads. 

The initiative stems from the company’s decision to unify its ML efforts into a central, efficient system capable of handling real-time use cases without driving up operational costs.

Meesho’s AI leadership emphasized that the platform was rigorously tested during its March 2025 Mega Blockbuster Sale, successfully scaling to handle surging traffic while improving engagement, conversions, and order volume.

“Great technology should scale impact, not just infrastructure,” Sanjeev Kumar, Meesho’s Founder and CTO, said in a statement.

“We put BharatMLStack to the test during high-traffic events like our Mega Blockbuster Sale in March 2025, where it delivered at scale—demonstrating its ability to perform under peak load conditions. This helped Meesho drive higher user engagement, better conversions, and increased order volumes during the sale. By open-sourcing it, we’re sharing a high-scale, AI stack with the broader tech community—purpose-built for real-time use cases and tailored for Indian businesses.”

Beyond raw performance, BharatMLStack is engineered to accommodate uniquely Indian internet behaviors such as transliterated search queries, low-resource devices, and fuzzy product discovery. Its modular components are built to support use cases like fraud detection, personalized search, ranking systems, and automated product tagging.

Debdoot Mukherjee, Meesho’s Chief Data Scientist, explained that the company initially relied on a mix of open-source and proprietary ML tools but found them either too expensive or inflexible.

“With proprietary platforms, it's like driving a sports car when all you need is a compact vehicle and you can't even look under the hood,” he told Moneycontrol. “Open source gives you transparency and control.”

Meesho's open-source release begins with its Online Feature Store, a tool that serves up precomputed features to ML models in real time. Such stores are crucial to ensuring that systems like product recommendations or fraud detection operate with low latency and high consistency, avoiding data silos across teams.

Empowering India’s AI builders

With this release, Meesho is aiming to equip India’s AI and data science community, including early-stage startups and independent ML engineers with tools that are not only scalable but also financially sustainable.

The company is not looking to monetize the release at this stage. Instead, the objective is to foster collaboration, invite feedback, and crowdsource improvements.

Mukherjee noted that early-stage startups without dedicated platform teams may benefit from BharatMLStack’s plug-and-play architecture, while more mature companies hitting scale could use it to build robust, cost-effective infrastructure without reinventing the wheel.

Additional components such as model serving infrastructure, model registration systems, and workflow authoring tools are expected to be open-sourced in phases over the coming months. Meesho says it will consider the initiative successful if even “a dozen companies” integrate the stack meaningfully into their core ML workflows.