Client engagement · anonymised
A Repeatable ML Deployment Platform for a French Professional-Services Firm
AI platform engineering for a French professional-services firm: Kubernetes, Terraform and CI/CD on GCP took deployments from days to minutes.
- Industry
- Professional services
- Region
- France
- Focus
- AI platform engineering
Client engagement — the client is anonymised at their request; figures are as measured on the engagement.
Challenge
A French professional-services firm of 10–50 staff had built several promising AI prototypes, but had no repeatable way to get them into production. Models were deployed by hand, there was no CI/CD, and every environment was configured differently — so each release took days and carried the risk of something behaving differently in production than in testing.
Solution
The approach: one AI platform on Google Cloud, with every ML workload containerised, every environment defined as code and every release flowing through the same automated pipeline. Instead of a bespoke deployment per prototype, each model now follows a single, standard path from commit to production.
Approach
- 1
We designed the cloud architecture on GCP around the firm's workloads, security needs and the prototypes it wanted to put into production.
- 2
We containerised the ML workloads with Docker so each model runs identically on a laptop, in testing and in production.
- 3
We built automated CI/CD pipelines in GitHub Actions to test, package and release every change.
- 4
We defined the infrastructure as code in Terraform, so every environment can be reviewed, rebuilt and kept consistent.
- 5
We set up a production deployment workflow on Kubernetes, with a clear path to promote, roll back and add new models.
Frequently asked questions
Why containerise ML workloads?
Containers package a model with everything it needs, so it behaves the same in development, testing and production. That removes the "it worked on my machine" failures behind many slow releases.
What does infrastructure as code change day to day?
Environments are described in Terraform and reviewed like code, so a change is visible before it is applied and any environment can be rebuilt identically.
Can new models use the platform without more engineering?
Yes. The pipeline and deployment workflow are shared, so a new model starts from the same template and follows the same release path.