LowPoly Shorts Engine — Prompt-to-Video AI Pipeline

Creating a polished vertical short involves repeatedly moving between ideation, image generation, narration, captions, motion, and editing. Many existing workflows are cloud-only and expensive to iterate with.

I built a single-prompt pipeline that creates a structured video plan, generates scene visuals, adds parallax motion, produces voiceover and word-level captions, renders a hook overlay, and exports the final 9:16 MP4.

The system uses a monorepo with FastAPI, Next.js, and a Python GPU worker. It supports deterministic prompt-and-seed runs, streams progress through server-sent events, and includes version-controlled QA reports, launch documentation, and a manual testing harness.

The engine has been tested end-to-end on a single host and optimized to run on a laptop-class 4 GB GPU. The core engine is working; the surrounding SaaS layer remains in closed beta.