European visual storytelling studio
Spatial canvas grammar interpreted by Claude on Bedrock, four agentic capabilities, sub-200ms canvas response.
Filmmakers and visual storytellers think in spatial and cinematic terms — scene composition, character blocking, camera movement — not in text prompts. The customer needed a platform that could interpret a spatial canvas as AI generation instructions. Nothing on the market could turn canvas-based spatial interactions into coordinated multi-model AI generation across depth estimation, pose detection, image generation, and Gaussian splatting.
A multi-agent AI Visual Storytelling Platform: (1) a Spatial Grammar Agent — Claude Sonnet on Bedrock interprets canvas element arrangements and decides generative actions; (2) a Scene Reconstruction Agent — depth estimation and Gaussian splatting on SageMaker turn 2D images into navigable pseudo-3D environments; (3) a Generative Director Agent — interprets OpenPose skeletons and triggers regeneration on g6e.12xlarge / p4de.24xlarge GPU instances; (4) Step Functions orchestration selects and chains models based on spatial context. ECS Fargate runs the canvas backend; DynamoDB persists spatial state.
Spatial grammar interpretation accuracy moved from 78% to 94% over the engagement. Canvas interaction response stays under 200ms via Lambda. Four agentic capabilities live in production: Magnetic Storyboard to Animatic, Image to Navigable Scene, Editable Generative Scene, and Drag-to-Direct AI filmmaking. Real-time multi-user collaboration via WebSocket. Artist-in-the-loop control held throughout via the spatial canvas interaction model.
Under the hood
Customer name redacted at the customer’s request. Numbers, services, and architecture are unchanged.
Next case study
AI creative tools startup