2024 · Platform
A self-hosted MLOps platform on Kubernetes
a ride-hailing company · Mlops engineer
The problem
A data team trained models on laptops with no shared tracking and no repeatable path from notebook to model.
What I built
I built the platform as code: a Kubernetes cluster with the standard open source tools for pipelines, notebooks and experiment tracking, joined behind one dashboard.
The result
Data scientists can go from a notebook to a tracked, managed model without leaving the company's cloud.
Main tools
- Amazon EKS
- Kubeflow
- MLflow
- Juju
- Terraform
This is the short version. The full case study is available on request.