Skip to content

Case: computer-vision product

A product that analyses movement in video: the user uploads a video and gets feedback generated by a language model. Described abstractly here ("App A").

flowchart LR
    ui["Web UI (SSR)"] --> api["API"]
    api --> proc["Video processor (GPU node)"]
    api --> llm["LLM gateway"]
    api --> db[("Database")]

Shape decisions

  • The UI talks to the API over the in-cluster service, never the public domain.
  • Heavy processing runs on a GPU node, called only by the API.
  • The API goes through an LLM gateway that centralizes keys, limits and cost observability.
  • Delivery follows the same flow as every app: PR, image, tag in the GitOps repository, sync.

What this case illustrates

GPU workload isolation, internal network boundaries and the GitOps pattern applied to a product with more than one component.