Production-grade MLOps on Google Cloud
Objective
Give Numan a way to take any machine learning model from a data scientist’s notebook to running safely in production.
Obstacle
Every model that made it past the experimentation stage had been deployed to Vertex AI by hand, one at a time, with no shared process behind it.
Outcome
A secure, fully automated MLOps platform on Google Cloud that any model can be shipped through.
Background
Numan is a UK platform built around men’s health, delivering clinician-backed treatment and ongoing care to customers online. As the business scaled, so did its reliance on data science: models to understand things like customer churn and treatment engagement were becoming central to how Numan made decisions. But building a model and running one in production were two different disciplines entirely, and Numan had no bridge between them.
Challenges
A Model Is Not A System: Numan’s data scientists could build a working model. Getting that model running automatically, every day, reliably, and safely, was a completely different skillset. Because there was no standard platform, each model that reached production had been set up by hand, individually. That meant no consistency, no easy way to check the setup was correct, and no way to confidently repeat the process for the next model.
Testing And Live Were Too Close Together: The space where new ideas got tried out sat too close to the systems the business actually depended on. There was no proper barrier stopping a mistake made while experimenting from reaching real customers and real decisions. As Numan built more models, the manual, one-off approach to deployment would only get riskier. They needed a platform that could take in any model and run it safely.
Solution
ML Platform: We created clearly separated spaces for testing, for running live, and for managing the whole system, so a mistake made while experimenting can never spill over into what the business relies on. The entire setup was built using code meaning it can be automatically rebuilt, checked, and audited at any time. Access was locked down throughout: no person can directly change the live system by hand, only automated, tightly controlled processes can.
CI/CD: We built an automated system that creates realistic test data and uses it to test the entire process from start to finish. When new code is approved and merged, the system automatically moves it forward. Once testing is complete, it cleans up everything it created. This helps us make sure all parts of the system are working correctly.
Proving It Worked: A platform is only worth something if it works under real conditions. So we took a number of Numan’s existing models around Lifetime Value and Churn, and put it through the new platform as the first live test. It now runs automatically every day, feeding the same reports the business already used, with monitoring in place to flag if its predictions started drifting. We also built in a permanent safety rule that applies to any model on the platform going forward: a new version is only ever allowed to take over if it’s proven to perform better than the one before it.
Impact
A Repeatable Path To Production: Numan no longer needs a bespoke, manual effort every time a model is ready to go live. The platform runs as three separate environments, development, control-plane, and production, so a mistake made while experimenting has no path into the systems the business depends on. Every part of it, from the GCP projects and networking down to the service accounts and permissions, is defined in code rather than clicked together by hand, so it can be rebuilt, checked, and audited at any time.
Workbenches: When data scientists work on models using their own individual setups, differences in software, tools, and configurations can cause problems. We created standardised workbenches for everyone, giving each data scientist the same tools, software versions, and setup. This makes it easier to collaborate, reproduce each other’s work, and move models into production without unexpected issues caused by differences in individual environments.
Safer By Design: No person holds direct access to production. Every deployment runs through a scoped service account via a CI/CD pipeline from version control, so changes ship without manual steps and without anyone needing production access at all. The churn model itself now runs automatically every day on Vertex Pipelines, feeding the same BigQuery outputs the business already relied on, with monitoring in place to catch drift early.
A Team That Owns It: Retraining is gated on quality: a new version of the model only ever goes live if it’s proven to outperform the one it’s replacing, so the system can improve but never quietly get worse. We handed over the full Terraform codebase and a runbook, so Numan’s team runs and extends the platform independently, with no ongoing dependency on us.
Creating Value For Numan...
1 platform built to run every future model,
100% output match before go-live.
And a scalable system for the future.
Get In Touch
Our friendly team are always on hand to answer questions, troubleshoot problems and point you in the right direction.
