video1.0<iframe src="https://www.loom.com/embed/588ed635099f4187acc398ca3ce095c2" frameborder="0" width="1668" height="1251" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>12511668Loomhttps://www.loom.com12511668https://cdn.loom.com/sessions/thumbnails/588ed635099f4187acc398ca3ce095c2-ef091114a70acf6c.gif766.889In this Loom, I walk through my screenplay instructor product with three parts, React front end, a FastAPI back end, and a retrieval pipeline. The UI sends analyze once to produce structure, then reuses that analysis and the current draft so every follow up gets re layered with the screenplay for instructor feedback, while keeping the API key off the browser. On the backend I parse scenes into metadata, embed chunks into ChromaDB using MiniLM on CPU, retrieve and rerank top k with a cross encoder. I ran three studies, error analysis, an ablation study, and a prompt comparison where full system analyst plus RAG performed best. No specific action was requested from viewers.