The transition addressed a critical bottleneck in Subaru’s workflow: massive 30 GB AI container images that previously required three hours to download. By implementing Envoy Gateway, Gateway API, and MetalLB, engineers reduced this duration to just three minutes. This change allowed the team to move away from manual scripts toward a more reliable, automated environment.

Ryoji Kobayashi, a DevOps engineer in the ADAS development department, noted that the primary goal was removing operational friction to let engineers prioritize model accuracy over infrastructure maintenance. The integration of Argo CD and Helmfile brought standardized GitOps practices to the team, while Argo Workflows automated the end-to-end machine learning pipeline. These technical improvements provided a scalable foundation that now supports more frequent iteration cycles for the next generation of Subaru’s EyeSight technology.