AI video production engine
Video Automation Service (VAS)
A system that turns a topic or script into a finished video through an AI editorial layer and a deterministic rendering pipeline.
Problem
Video generation becomes unreliable when creative choices, media assets, rendering, and external providers are mixed into one opaque request.
Constraints
- Slow and failure-prone external providers
- Expensive rendering
- Partial reruns should not repeat completed work
- Editorial output needs structure
Approach
AI produces a typed editorial and edit-plan contract. A FastAPI control plane and workers resolve assets, render deterministic scenes with Remotion and FFmpeg, and retain the state needed to resume safely.
Architecture
Key decisions
Plan as a contract
A structured edit plan separates editorial reasoning from rendering execution.
Cache at the asset boundary
Completed clips and assets can be reused, avoiding an expensive full regeneration after a small edit.
Fallbacks are explicit
Provider routing is a system decision with cost and capability boundaries, not a hidden prompt trick.
Failure handling
Jobs persist their stage and outputs. Workers retry safe actions, provider fallbacks can be selected, and failed rendering stays recoverable rather than becoming a blank final result.
Result
A production-scale implementation with substantial automated coverage and multiple provider integrations, built to make a complex pipeline observable.
What I’d change next
Move more rendering capacity to remote workers while preserving the same recovery model.