Our client has sensors deployed at the edge across many locations, with images being returned to the cloud for inference. They are currently relying on computer vision models developed using AutoML, and deployed via Vertex AI. The accuracy of the AutoML models was not sufficient for their use case. Further, the compute costs and lack of flexibility of using Vertex AI, along with the fragmented data science workflow, were prohibitive as the company began to scale. They needed a more streamlined, accurate and cost effective approach. Finally, the system needed to offer high availability, and self healing properties to enable data scientists to manage the platform without relying on an engineering team.

We implemented a multi-stage computer vision pipeline to process sensor images and detect objects of interest using a custom trained machine learning model. While our initial solution was deployed using VertexAI, it became clear that a more customizable and cost effective approach was required. Toward that end, our team migrated the pipeline to Kubeflow, allowing their data science team to streamline the development and experimentation of pipelines and seamlessly deploy a prototype pipeline into production. This new approach will be able to scale with our client and facilitate the deployment of additional solutions as their company grows.