How an AI webinar builder chains script, slides, TTS and a Q&A host into one pipeline

I’ve been looking at EverPitch, an AI tool that turns a single input into a finished, self-running webinar. I’m not a user yet. I studied the demo for a review, and the architecture caught my eye more than the marketing did.

As far as I can tell from the demo, the pipeline works like this:

Input normalization. It takes a one-line prompt, a guided form, a PDF, a URL or an old recording, and turns each one into the same source material.
Script generation. A slide-by-slide script is written, so every slide gets its own spoken text. That makes partial regeneration cheap.
Layout mapping. The script is fitted into 73 slide templates across 11 visual styles.
TTS. 30 voices across 87 languages, plus voice cloning.
Scheduled playback. Sessions run in the viewer’s own time zone, with registration, reminder emails and checkout (Stripe, PayPal and others).
Live Q&A. An AI host answers chat questions using only the webinar’s own content. My guess is retrieval over the script, so it stays on topic instead of making things up.

A few design choices I found interesting:

Prompt-based editing with a before/after diff. It’s basically a review step before you commit a change.
Grounding answers in the content. Limiting the Q&A host to the webinar’s own material is a sensible way to keep it from inventing claims in front of buyers.
Credit-based pricing. You pay credits for generation. The vendor doesn’t publish the cost per webinar, which I’d like to see.
Hard 60-minute cap per webinar. That’s probably a TTS and render cost limit.

There’s a free live reveal on Nov 15 (8 PM EST) and Nov 16 (10 AM EST) if you want to see the pipeline running. I wrote up the details here

Curious how others would build the Q&A part. Would you use RAG over the script, or fine-tune per webinar?