Meesho open-sources in-house ML platform BharatMLStack to help startups scale AI faster

Scale, stress-tested

In FY25, BharatMLStack processed an average of 1.91 petabytes of data per day, handled 66.9 trillion feature retrievals, and supported 3.12 trillion real-time inferences at peak.

“Millions of requests per second, trillions of features being served daily – there will be very few systems in India, or even globally, that operate at this scale,” Mukherjee said.

Unlike typical enterprise platforms that scale by “throwing boxes” at the problem, Meesho had to optimise for cost and reliability – critical in India’s low-margin e-commerce context. The stack was built to deliver not just performance, but resilience, especially during high-traffic bursts like monthly sale events.

“Of course you can scale by throwing boxes at the problem, but you’ll end up spending much more than you can afford. If your infrastructure isn’t efficient, it directly impacts your profitability,” Mukherjee said. “That’s been the hallmark of this project — building something that’s extremely stable and cost-efficient even under peak loads.”

Beyond traffic handling, the stack powers a wide array of machine learning use cases at Meesho, including ranking and recommendation systems, personalised search, fraud detection, and automated product tagging. It’s also built to handle uniquely Indian behaviours, such as transliterated search queries and fuzzy product discovery.

Why now, and why open?

Meesho is not looking to monetise the release at this stage. Instead, the move is part of a broader attempt to support India’s growing ML developer base, and to gather feedback and contributions that could improve the stack further.

“We tried a bunch of open-source and proprietary tools. Many of them proved too costly, and the rest lacked flexibility,” Mukherjee said. “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. Open source gives you that transparency and control.”

The current release, the feature store, helps machine learning models retrieve pre-computed data points in real time. That allows for faster, more accurate predictions while simplifying infrastructure and avoiding data silos.

Mukherjee believes the stack could be particularly useful for mid-stage startups starting to hit scale. “Early-stage startups may not have platform teams, so a plug-and-play solution like this could help. But I’d especially recommend it to startups that are going big on ML and want something scalable and cost-efficient,” he said.

More to come

Meesho plans to open-source other parts of the stack gradually. including model serving infrastructure, model registration systems, and workflow authoring tools.

Mukherjee said the company would consider the effort a success if even “a dozen companies start using it very effectively to drive their core ML products.”

“We’re already investing in this platform internally to run Meesho’s business. But if people start adopting it and contributing, we’ll get feedback that helps make it more complete, robust, and efficient. That’s something we’d really value,” he concluded.