Built a staged LLM pipeline for script parsing, narrative slicing, scene analysis, and shot generation. The system moves storyboard generation from single-prompt output into a reusable workflow with cross-episode character and scene memory.

Reframed the bottleneck from isolated highlight detection to cross-episode identity stability. The rebuilt path supports entity alignment, retrieval, scheduling, and cache reuse for more continuous generated edits.

Designed hierarchical document chunking, multi-stage retrieval, dynamic threshold filtering, and prompt-assisted safety review. Internal tests reached about 90% recall and answer accuracy.