How it is made
Two AI hosts, Paul and Jane, make each episode with an editor: you, through your AI assistant.
- You ask your assistant for an episode. It searches the sources itself, and checks the questions listeners recorded on this page.
- A writer agent drafts the script from those sources only, citing them line by line, with chapters and a summary. It looks up what the hosts said in past episodes, so they refer back instead of repeating themselves.
- You review it in the conversation. Your assistant compares each line with its sources; you ask for changes until you approve.
- The hosts' voices record it, one voice per host and emotion, and the episode lands in the podcast feed, with its transcript and chapters.
No assistant at hand? Make an episode in the browser: the hosts search the web for their sources, or read the links you give them, and you review the script the same way.
Made with
- Symfony AI: the writer agent and its tools, the structured script, the show's memory (vector store), and the MCP server your assistant talks to.
- Mistral, for everything the hosts do: writing (Mistral Medium), speaking and listening (Voxtral), remembering (Mistral Embed).
- Symfony 8, Symfony UX, FrankenPHP and PostgreSQL with pgvector.
Given back
Building this show led to two contributions to Symfony AI: Voxtral text to speech in the Mistral bridge, and a fix to the Postgres store's hybrid search.
Built for the Symfony AI & MCP Hackathon, November 2026.