Rebuilding a model release pipeline to ship in hours

Context

The client builds and ships machine learning models, and getting a new model into production took 3 weeks.

Client details are withheld for confidentiality. The facts below are limited to what we can verify.

The challenge

A slow release process means a finished model waits a long time before it does any useful work, and each release carries more change and more risk.

The client needed a release pipeline that could take a model to production quickly and repeatably.

What we did

1. 1

   Map the release path

   We traced how a model moved from training to production and found the steps that slowed it down.

2. 2

   Automate build and packaging

   Model packaging and checks moved into an automated pipeline, so the same steps run the same way every time.

3. 3

   Deploy through Git

   Releases were driven from version control, so every deployment is recorded and can be rolled back.

4. 4

   Hand over to the client's team

   We documented the pipeline so the client's own engineers can run and change it.

Results

• Model deployments cut from 3 weeks to 4 hours
• A repeatable release process instead of manual steps
• Every deployment recorded in version control