SnapMedia: an AI document to video pipeline

SnapMedia turns a document into a narrated video. You upload a PDF, and what comes out the other side is a finished, motion animated film with generated voiceover.

GPT-4o Vision handles extraction, pulling structured data out of the uploaded PDF rather than treating it as a wall of text. Vision models are good at this and they fail in a specific way that matters: when they get something wrong, they get it wrong confidently and quietly. Nothing about the output looks broken.

That failure mode is why the review editor exists, and it is the actual product. Extracted data lands in a React editor where a person checks it and corrects anything wrong before a single frame renders. No output reaches a customer without a human having looked at it.

Speed without that gate would be a liability rather than a feature. Any pipeline that puts a model's raw output in front of an end user is one hallucination away from an incident, and the only reliable fix is to make verification a required step rather than an optional one.

Once the data is confirmed, a Remotion pipeline builds the video programmatically: motion animated scenes composed from the verified content, with narration generated through OpenAI's Nova voice, then rendered out to a finished file.

Next.js and TypeScript throughout, GPT-4o Vision for extraction, Remotion for the render pipeline, OpenAI text to speech for narration.