P-001 · Case study
SHIPPEDVideo Generation Pipeline
Idea in, finished cinematic video out. Multi-phase agentic pipeline with per-atom review gates and cross-family LLM-as-judge. Millions of impressions a month.
Case study
The problem
JuiceLabs needed a pipeline that could take an idea and produce a finished cinematic video ad — not a demo, but production output shipping to enterprise clients at scale. Nothing off-the-shelf existed that handled the full journey from brief to final video with quality gates at every step.
The approach
Multi-phase agentic architecture. The brief turns into multiple storyboard candidates via parallel insight extraction and brainstorms. The user picks one, the system breaks it into shots, generates character reference frames, then composes the final video via a reference-to-video model. Multi-model orchestration across Opus 4.6, gpt-image-2, Seedream 4.5, and Seedance 2.0. A different model family judges the output so the generator and grader don't collude. Discrete user-approval gates at every expensive step.
What worked
Pipeline is live and shipping cinematic ads at millions of impressions a month for enterprise clients. Length-agnostic architecture means it can scale to any format without rearchitecting.
What I'd do differently
The cross-family judge approach was critical. Early versions where the same model family both generated and evaluated drifted into self-reinforcing quality bubbles. Using a different family for evaluation broke that loop.
More detail
Multi-phase agentic architecture: brief turns into multiple storyboard candidates via parallel insight extraction and brainstorms, you pick one, the system breaks it into shots, generates character reference frames, then composes the final video via a reference-to-video model. Multi-model orchestration (Opus 4.6, gpt-image-2, Seedream 4.5, Seedance 2.0). A different model family judges the output so the generator and grader don't collude. Length-agnostic. Currently shipping ads at millions of impressions a month for enterprise clients.