Building a platform from scratch is about standardization and developer experience.

The Problem

Every data science team was deploying models differently. Some used Flask on EC2, others used Lambda, and some were just running notebooks manually.

The Solution

We built a centralized platform on top of Databricks and AWS.

Key Components

  1. Feature Store: Centralized feature definitions using Feast.
  2. Model Registry: MLflow for versioning and stage management.
  3. Deployment Pipelines: GitHub Actions + Terraform to provision serving infrastructure automatically.
module "inference_service" {
  source = "./modules/inference"
  
  model_name = "fraud-detection"
  version    = "v2.1.0"
  min_instances = 2
  max_instances = 10
}

This reduced the “time-to-production” from weeks to hours.