Case study · FMCG · MLOps

Standardizing FMCG ML Operations with MLOps Best Practices Across Multiple Clouds

A global FMCG/CPG company struggled to scale ML use cases across several cloud platforms. DS Stream applied MLOps best practices to build a cloud-agnostic tool and a standardized MLOps template for GCP, Azure and Databricks. The result was streamlined maintenance, lower overhead costs, a faster path from development to production and better resource utilization.

Talk to our team See all case studies

Results at a glance

Maintenance tasks were streamlined, which cut overhead costs and raised operational efficiency.
CI/CD pipelines built with GitHub Actions accelerated the development cycle and moved models to production faster.
A Q&A chatbot powered by ChatGPT and Langchain gives data scientists real-time guidance, boosting satisfaction and productivity.
ML training and inference now run on one standardized, scalable setup across GCP, Azure and Databricks.
The project drove cost savings and improved resource utilization, so the client can scale ML use cases more effectively.

Why was the client struggling to scale ML use cases?

The client ran ML use cases on more than one cloud platform and found it hard to scale and expand them. There was no common way of working, so every environment needed its own effort. The company needed one standard for ML operations that kept training and inference simple and stayed flexible across environments.

What the client asked for

Standardized ML operations across cloud platforms
Streamlined ML training and inference tasks
Flexibility and scalability in every environment

What did DS Stream build to standardize MLOps across multiple cloud platforms?

DS Stream built two core assets: a cloud-agnostic tool and a customized, standardized MLOps template. Together they integrate the client's ML use cases across GCP, Azure and Databricks. The template packages MLOps best practices, so teams start every new use case from the same proven structure instead of building pipelines from scratch.

Key components

Automated CI/CD processes aligned with the client's organizational standards, implemented with GitHub Actions
Resource autoscaling and consumption logging to optimize cloud resources
A Q&A chatbot built on ChatGPT and Langchain that helps data scientists navigate the tool and templates in real time

How does the template keep models reliable after deployment?

The FMCG MLOps template includes model versioning, monitoring and retraining pipelines. Retraining is triggered automatically when the system detects data drift or concept drift. This means models are refreshed when the data or the patterns behind them change, not on a fixed schedule or after someone notices a drop in quality.

Versioning: every model version is tracked
Monitoring: model behavior is observed in production
Retraining: pipelines start on detected data or concept drift

Technology stack behind the solution

The solution combines cloud platforms, container tooling and generative AI. All of it serves one goal: the same MLOps workflow on every platform.

Cloud and data platforms: Azure, GCP, Databricks
Engineering and runtime: Python, Docker, Kubernetes
Automation: CI/CD with GitHub Actions
AI assistant: ChatGPT, Langchain

Need a similar setup? Explore our MLOps implementation and machine learning pipeline services.

Questions about this MLOps case study

Is the MLOps template tied to one cloud provider?

No. DS Stream designed the tool to be cloud-agnostic. The customized MLOps template integrates ML use cases across GCP, Azure and Databricks, so the client can run training and inference on the platform that fits each use case while keeping one standard way of working.

What triggers model retraining in this solution?

Retraining pipelines start when the system detects data drift or concept drift. They work together with model versioning and monitoring, so a new model version is trained when the input data or the underlying patterns change. This keeps models relevant without relying only on manual checks.

How do data scientists learn to use the tool and templates?

DS Stream deployed a Q&A chatbot powered by ChatGPT and Langchain. It offers real-time support for data scientists as they navigate the cloud-agnostic tool and the MLOps templates. According to the case study, this support boosted user satisfaction and productivity.