AI Video — Video Understanding + Generation Foundation

• Area: Intelligence
• Path: services/ai/video
• Kind: Video understanding (transcript + scenes + keyframes + summary) and generation (proxy)
• Status: v0.1.0 — foundation shipped (2026-05-24); async JobRun store, MP4 ftypmoov/mdat stub, pipeline allow-list, Go SDK

Role in the stack

video is the single integration point for both video understanding and generation. Understanding orchestrates voice (audio track) + imaging (frame analysis) + scene detection — pure composition, no new model code. Generation is proxy-only in v1 (Sora, Veo, Runway, Pika via services/ai/gateway); self-hosted text-to-video stays out of scope until open models reach G4 capability.

It is the Koder analog of a video-stack assemblage that today requires pasting together OpenAI Sora + Whisper + GPT-4V + custom scene-detect — the foundation collapses that into a single normalized API with one auth boundary, one quota system, one billing event stream.

Boundary vs neighbors

• services/ai/voice is the audio sibling and a hard dependency for analyze.
• services/ai/imaging is the frame sibling for caption/keyframe.
• services/ai/vision is the single-image counterpart.
• `products/media