Experimenting with conversational interfaces for long-form documents

[Pushkal Shukla]

Hi everyone,

I’m Pushkal, a senior backend ML engineer working on a small side project called PaperPod.

I’ve been experimenting with a simple idea:

Instead of asking an LLM to summarize a long document, what if two AI hosts could have a conversation about the document — and you could interrupt them and ask questions whenever something caught your attention?

I ended up building a working system around this.

It takes PDFs, research papers, notes, etc., generates a two-host dialogue, synthesizes the conversation into audio, and keeps the transcript synchronized with playback. During the conversation, you can also ask questions either from the document itself or using the document + web as context.

Interestingly, most of the engineering work wasn't in the basic LLM call. I spent much more time on routing, token reduction, prompt caching, TPM-aware scheduling, retries, TTS concurrency, and making long/diagram-heavy documents behave reliably.

As a recent test, an 18-page research paper with architecture diagrams and tables became a ~25-minutes conversation in ~70 seconds.

I’m curious whether this interaction model is actually useful beyond being a more entertaining form of summarization.

If anyone here works on document intelligence, RAG, voice AI, or LLM applications, I'd genuinely appreciate your perspective — particularly where you think this interaction model could work well, and where it probably doesn't.

I’ve put the experiment here if anyone wants to try it:
Try PaperPod

Thanks,
Pushkal