Our client has sensors deployed across many locations with scaling plans to reach the 100,000s of sensors in the next 12 months. Their data science team has developed promising models to detect outlier behavior based on the time series. However, these models were built in Jupyter notebooks, and required a plan and architecture to achieve a production system. The client needed an approach that was considerate of their team’s current capabilities (they lacked strong ML Engineering expertise) and compute costs as the company begins to scale. They needed a near term production deployment and a longer term roadmap to support growth.
We designed and implemented a platform to support the full MLOps lifecycle. We partnered with our client's data scientists to refactor their notebook code into reusable modules, while simultaneously providing mentorship on software engineering best practices. Based on their current ML needs, with model deployments automated by CI/CD processes, AWS Sagemaker Batch Inference was an ideal, low-cost solution to serving their models in production. Eventually, as they scale the company and expand their data science practice, an open source Kubernetes-based platform such as Kubeflow will provide a more cost-effective approach, countered by an increased need for ML engineering expertise on their team.