Meesho Set to Open Source its ML Platform

At the Nvidia Summit in Mumbai, eCommerce giant Meesho announced that it is open-sourcing its ML platform.

Debdoot Mukherjee, head of AI at Meesho, said, “We will be open-sourcing our ML platform, including our feature store and model orchestrator. We believe this move will empower developers and data scientists everywhere to leverage the power of AI more effectively.”

In a presentation titled “Cost Effective High-Performance Model Serving for the Masses,” Debdoot presented a deep dive into Meesho’s ‘massive scale’ machine learning models. 

He spoke about Feature Store, an architecture designed for high-performance machine learning operations. The architecture helps Meesho standardise and manage features for their machine learning models at scale, bridging raw data and their ML models. 

The presentation also revealed that this system could handle millions of requests every second, and 99 per cent of these requests are handled in under 10 milliseconds. Debdoot added that this Feature Store architecture is also set to be open-sourced.

In a LinkedIn post, Debdoot said, “We focused on our journey of scaling transformer-based ranking models in production, highlighting the optimizations we’ve implemented with TensorRT to ensure we meet stringent latency and cost requirements.”

A few days ago at Build With Meta Summit,  Meta revealed that Meesho was among many other consumer apps planning to implement Meta’s open-source Llama models inside their tech stack. 

Over the last few years, Meesho has proactively integrated generative AI, and machine learning capabilities into its platform. 

Internally, their large-scale decision engine has incorporated several features to enhance code maintainability, reduce system latency, and provide an improved error-handling system with multiple fallback solutions. Further, the decision engine also helps developers monitor system health by creating a dashboard to track and spot problems easily. 

On the user front, Meesho uses generative AI to simplify the listing process for sellers within the platform. They’ve also employed features to translate product descriptions, addresses, and voice searches into other Indian languages besides Hindi. Meesho also uses Microsoft’s Azure OpenAI to optimise call centre costs and operations, reportedly resulting in a 20% boost in customer effort scores. 

Last year, Meesho partnered with IISc to advance their research in generative AI and develop multimodal representation learning and other generative AI capabilities to enhance the user experience. 

In a conversation with AIM last year, Debdoot said, "At Meesho, AI and ML are fundamental components to solve a wide range of problems across every aspect of an e-commerce platform.” 

Meesho is one of India’s fastest-growing e-commerce platforms, with around 120 million active users. Open-sourcing, a machine learning platform with proven capabilities in handling high-throughput feature serving, may benefit the ecosystem, help developers build similar ML systems for their platforms, and solve existing challenges in their technology.