Our client has a medical device capable of differentiating cancer tissues and informing precision medicine treatment decisions. The client had high performing models in Jupyter notebooks and needed a path to production deployment on their individual devices deployed inside hospital networks. The devices need to be able to run in offline environments, and the model must make real-time decisions that impact device measurements. With interest in deploying multiple models, our client needed the ability to monitor model versions and performance across multiple devices in order to conduct timely retraining processes and improve model capabilities.

We productized our client’s models via an API endpoint that interfaced with their existing data systems to trigger real-time model predictions upon data upload from their device. The API endpoint is supported by a combined MLFlow and Docker strategy that enables automated deployments and version control of production models. Additionally, we employed MLFlow during the model development phase to track and document model performance and maintain a version controlled model registry. Our approach will facilitate the scaling of complex machine workflows and applications while minimizing the manual load associated with model deployment, monitoring and retraining.