meesho/bharatmlstack

BharatMLStack is an open source machine learning infrastructure platform built by the Indian e-commerce company Meesho to run large scale ML workloads, both in real time and in scheduled batches. It is meant to run on any cloud provider, on premises, or at the edge, so a company is not locked into one vendor, and it is built to run on Kubernetes. The stack is made up of several separate components that work together. An online feature store serves the numeric and categorical inputs a model needs in under 10 milliseconds, even at millions of queries per second. Inferflow orchestrates real time model inference as a series of connected steps. Numerix is a math engine written in Rust for fast matrix calculations. Skye handles vector similarity search with support for different backend databases. There is also an interaction store for logging user behavior signals, a control plane called Horizon that coordinates all the services, and a web console called TruffleBox for managing and approving features. Go and Python client libraries are provided so applications can talk to these services easily. According to the README, the platform has been used in production at Meesho to power things like personalized product recommendations, ranking of search results, fraud detection, image based search, and large language model powered recommendation systems. It reports handling over a million queries per second for model inference and hundreds of thousands of queries per second for embedding search, with reported uptime above 99.99 percent. Getting started involves cloning the repository, setting version numbers for each component, and running a start script that uses Docker Compose, with a separate quick start guide covering sample data and health checks. The project welcomes outside contributions through a documented contributing guide and has a Discord community for support. Its source code is released under the BharatMLStack Business Source License, a source available license rather than a fully permissive open source one, so anyone evaluating it for commercial use should read the license terms carefully.