BharatMLStack is a production-ready, cloud-agnostic ML infrastructure platform that powers real-time feature serving, model inference, and embedding search at massive scale. Built and battle-tested at Meesho, it is designed to help organizations ship ML to production faster, cheaper, and more reliably.
BharatMLStack is built around four core tenets:
Ship ML to production faster than ever.
- 3x faster experiment-to-deployment cycles
- 95% reduction in model onboarding time
Run anywhere. Own your stack.
- Runs across public cloud, on-prem, and edge
- Kubernetes-native with zero vendor lock-in
Do more with less.
- 60–70% lower infrastructure costs vs hyperscaler managed services
- Optimized resource utilization across CPU and GPU workloads
Enterprise-grade reliability at internet scale.
- 99.99% uptime across clusters
- 1M+ QPS with low latency
Built for the demands of one of the world's largest e-commerce platforms:
| Metric | Performance |
|---|---|
| Feature Store | 2.4M QPS (batch of 100 id lookups) |
| Model Inference | 1M+ QPS |
| Embedding Search | 500K QPS |
| Feature Retrieval Latency | Sub-10ms |
| Component | Description | Version | Docs |
|---|---|---|---|
| TruffleBox UI | Web console for feature registry, cataloging, and approval workflows | v1.3.0 |
Docs |
| Online Feature Store | Sub-10ms feature retrieval at millions of QPS with streaming ingestion | v1.2.0 |
Docs |
| Inferflow | DAG-based real-time inference orchestration for composable ML pipelines | v1.0.0 |
Docs |
| Numerix | Rust-powered math compute engine for high-performance matrix ops | v1.0.0 |
Docs |
| Skye | Vector similarity search with pluggable backends | v1.0.0 |
Docs |
| Go SDK | Go client for Feature Store, Interaction Store, and logging | v1.3.0 |
Docs |
| Python SDK | Python client libraries for Feature Store and inference logging | v1.0.1 |
Docs |
| Interaction Store | ScyllaDB-backed store for user interaction signals at sub-10ms | — | — |
| Horizon | Control plane that orchestrates all services and powers TruffleBox UI | v1.3.0 |
— |
Full documentation at meesho.github.io/BharatMLStack | Blogs
git clone https://github.com/Meesho/BharatMLStack.git cd BharatMLStack/quick-start #Set versions ONFS_VERSION=v1.2.0 HORIZON_VERSION=v1.3.0 TRUFFLEBOX_VERSION=v1.3.0 NUMERIX_VERSION=v1.0.0 ./start.sh
For step-by-step setup, Docker Compose details, sample data, and health checks, see the full Quick Start Guide →.
BharatMLStack powers a wide range of ML-driven applications:
| Use-Case | What BharatMLStack Enables |
|---|---|
| Personalized Candidate Generation | Retrieve and rank millions of candidates in real time using feature vectors and embedding similarity |
| Personalized Ranking | Serve user, item, and context features at ultra-low latency to power real-time ranking models |
| Fraud & Risk Detection | Stream interaction signals and features to detect anomalies and fraudulent patterns in milliseconds |
| Image Search | Run embedding search at 500K QPS to match visual queries against massive product catalogs |
| LLM Recommender Systems | Orchestrate LLM inference pipelines with feature enrichment for next-gen recommendation engines |
| DL & LLM Deployments at Scale | Deploy and scale deep learning and large language models across GPU clusters with Inferflow orchestration |
We welcome contributions from the community! Please see our Contributing Guide for details on how to get started.
- Discord: Join our community chat
- Issues: Report bugs and request features on GitHub Issues
- Email: Contact us at [email protected]
BharatMLStack is open-source software licensed under the BharatMLStack Business Source License 1.1.