AI & MLLLMOpenAIDjangoNode.jsLivekit

Building an Audio Performance Application Around AI Generated Scripts

A generated story becomes more interesting when people can perform it. For a project that supports live audio performances, I brought that idea into an application: create a structured script, assign roles, and support the surrounding performance workflow.

The project spans an earlier Django implementation and a later Next.js application. Both connect language-model generation to product requirements rather than treating a block of generated prose as the finished result.

The script has to work for performers

A performance script needs more structure than a story summary. Characters must map to the cast, dialogue must be attributable, scene descriptions must be readable aloud, and the result must fit the application's length and format constraints.

That makes the output contract central. The model generates material that application code has to parse, validate, display, and hand to participants.

The retained implementations include OpenAI and Anthropic provider integrations. Provider selection is part of the generation system, while validation remains an application responsibility.

Several AI capabilities in one product

The Node version also contains a generated-poster pipeline. A poster and a script are different artifacts, even when they share a show concept. Each needs its own input preparation, output handling, and review.

LiveKit provides the real-time audio layer for performers. That infrastructure should be understood separately from model generation: using an audio room does not mean the performers' voices are generated by AI.

This division makes the product easier to explain. Language models help create material; conventional application code manages its structure and workflow; real-time media infrastructure connects people.

From an integration to an application

The retained Django history begins in September 2025, and the Node version has later development history. Those versions belong to the same product story rather than representing unrelated AI experiments.

The engineering questions grow as the workflow becomes more complete. How does the application handle invalid output? What happens when generation fails partway through? How are a script's cast and the participating users kept consistent?

The separate script-quality workstream addresses another important question: a valid script can still be uninteresting to perform. Structure and creative quality need different kinds of review.

What the project demonstrates

This project is useful in my AI history because it combines generation with a demanding output contract and a human activity. The model's response becomes one component in a larger application.

The implementation and retained documentation support that architectural account. I would evaluate public availability separately rather than treating local production notes as proof that a particular deployment is currently live.

A useful public demonstration would show a synthetic show brief, a short approved script excerpt, and the corresponding cast view. It should keep user records, operational configuration, and unpublished material out of the example.