MaranataClub: an end-to-end generative AI pipeline

Context

Audio content is easy to publish and hard to learn from. Forty-two episodes of Christian discipleship existed; what did not exist was anything that turned listening into study.

Problem

Turning audio into a learning experience means transcription, study material, assessment and illustration: four different generative problems, each of which is a project on its own if done by hand across forty-two episodes.

Approach

I built a complete pipeline: Whisper for transcription, GPT for study material and quizzes, and DALL·E 3 and Leonardo AI for automated visual generation. On top of it sits Immersive Mode, which synchronises the transcript and the AI-generated illustrations with the audio in real time. The platform runs on React 18, Firebase and Cloud Functions with a Feature-Sliced Design architecture and end-to-end tests, and adds passwordless authentication by magic link, collaboration with granular permissions and community voting.

Non-obvious decisions

• Generate once, at ingest, and store the artefacts. Transcript, quizzes and illustrations are deterministic outputs of an episode, so regenerating them per listener would pay the model cost forever.
• Synchronise the illustrations to the audio rather than to the page. Immersive Mode's whole value is the alignment; a gallery beside a player is a different and weaker product.
• Passwordless by magic link. The audience is not technical and the content is not sensitive, so a password is a support cost with no compensating benefit here.
• Feature-Sliced Design with end-to-end tests from the start. A generative pipeline has many moving parts and the failure that matters is the integration, not the unit.

Outcome

A full-stack product where generative AI is the production line rather than a feature: 42 episodes with synchronised transcripts, AI illustrations, study material and quizzes, plus collaboration with granular permissions and community voting.