Sebastian Menke, Head of Clinical AI, Savana
At Savana, we continuously advance clinical Natural Language Processing (cNLP) by developing high-performance models that drive our applications. To maintain production-quality standards and streamline our MLOps processes, we utilise ClearML, a powerful AI/machine learning platform that enhances our model development and deployment efficiency. Here’s how ClearML has transformed our workflows.
High-throughput cNLP model quality checks with ClearML pipelines
One of our key challenges is ensuring that our cNLP models perform reliably in production environments. By using ClearML Pipelines, we have built a highly modular and configurable processing pipeline that mirrors our production high-throughput processing pipeline. This allows us to run production-like quality checks before deploying models, ensuring robustness and performance consistency.
Modular and scalable inference execution
Our cNLP pipeline needs to be flexible enough to handle diverse hardware and software requirements. ClearML provides us with:
- Modular pipeline configuration: We can adjust pipeline behaviour using multiple command-line arguments, making it easy to test various configurations.
- Scalability: Whether it is running on GPUs, CPUs, or different software environments, ClearML enables seamless model inference execution across different infrastructure setups.
User-friendly UI for efficient experiment tracking and execution
With ClearML’s intuitive UI, our data scientists can:
- Monitor pipeline executions in real time
- Relaunch pipelines with pre-defined parameters
- Adjust pipeline settings on the fly – all without needing additional coding
This self-service capability significantly speeds up experimentation and iteration cycles, allowing us to optimise models more effectively.
Transparent logging and metadata storage for compliance
Maintaining compliance and procedural documentation is critical. ClearML helps us achieve this by:
- Automatically storing logs and metadata for each pipeline execution
- Ensuring we meet regulatory and procedural requirements through full traceability
This feature is crucial for auditability and for keeping a well-documented record of all our MLOps activities.
Flexible Jupyter sessions for experimentation
Beyond pipelines, our team relies on ClearML’s Jupyter Sessions as a daily workstation. This feature allows us to:
- Run code on custom hardware and software settings based on specific use case requirements, whether the hardware is on-site or in the cloud
- Experiment with different models and configurations flexibly, without the overhead of manually managing resources and environments
Empowering data scientists with MLOps autonomy
By establishing a robust permission matrix using ClearML’ RBAC, we have enabled us to have granular permissions, allowing our data scientists to access compute resources and projects – without constant DevOps support. This autonomy accelerates our workflows and frees up DevOps resources for higher-level infrastructure tasks.
Conclusion
ClearML has become an essential component of our AI building ecosystem at Savana. By leveraging its pipelines, UI, logging, and flexible Jupyter environments, we have optimized our cNLP workflows, improved compliance, and empowered our data scientists. The result? Faster experimentation, more reliable models, and a smoother path to production.
Editor’s Note: To see ClearML in action, please request a demo.