Public cloud couldn’t meet our shop-floor constraints: real-time Vision AI requires immediate decisions within strict manufacturing cycles, and factory video feeds must stay on-prem for data sovereignty.

At Ford Otosan, we run a multi-site K8s cluster with a centralized control plane and remote GPU nodes distributed across plants over a private network. By moving inference to our on-prem edge datacenters, we achieve the low-latency performance required for real-time intervention.

In this talk, we explore:

Architecture: Managing multi-site K8s and handling site-link failures.

GPU Strategy: Scaling inference (Triton/vLLM) with HPA, and optimizing sharing via MIG/Time-slicing.

Workload Segregation: Separating strictly on-prem serving from cloud-eligible batch pipelines.

This case study provides a blueprint for leveraging open-source tools to solve data privacy and latency challenges at the Edge.