🪴 Brain

Travelcast AI

3 min read

Travelcast AI

🗒️ Description

An AI-powered travel-podcast platform in two repos: a generator (Python/FastAPI) that researches a destination with parallel AI agents, synthesizes a script, and produces multilingual audio; and a website (Next.js) that publishes, archives, and streams the finished episodes. The generator publishes to the website over an authenticated REST API.

🧩 How the two repos relate

  • trip-podcast-generator — research → synthesis → audio. The producer.
  • trip-podcast-website-cursor — public multilingual portal. The consumer. Episodes land at public/episodes/{slug}/{language}/ as metadata.json + podcast_final.mp3 + notes.md (file-based, no DB on the site).

Technology

Stack — Generator

  • Python 3.11+, FastAPI, Celery on Redis for the task queue, Supabase (Postgres) for place/episode state.
  • Research: Google Maps/Places API (POIs), Perplexity API (facts), web scraping for vibe/critic content.
  • LLM: OpenAI GPT-5.5 for consensus synthesis and translation.
  • TTS (multi-provider fallback): ElevenLabs (primary, multilingual) → OpenAI TTS → Chatterbox (local, torch-based, free).
  • Audio: FFmpeg normalization (stereo, 44.1 kHz, 192 kbps).

Stack — Website

⚙️ Architecture know-how (reusable)

  • Parallel research, single synthesis. A Celery chord runs four agents concurrently for the English master — Geo (Maps), Historian (Perplexity), Vibe (blogs/forums), Critic (reviews) — then one GPT consensus pass synthesizes them without filtering to preserve nuance. Translations reuse the English script; no re-research per language.
  • Provider-factory TTS with per-episode fallback. Audio stays consistent within an episode (one provider chosen up front, after a credit preflight) but flexes across episodes. New providers implement one ABC; voice IDs + provider chains live in a central languages.json, not in env.
  • Config-file-driven languages. Adding a language is a config edit (voice IDs, provider order), not a code change. A pronunciation agent handles proper nouns from a config override file.
  • File-native publishing. The site’s truth lives on disk per episode/language; the generator POSTs to it with a timing-safe bearer-token comparison.
  • Queue resilience: task_acks_late, reject-on-worker-lost, a 2-hour visibility timeout, and explicit orphan-task detection in API endpoints to recover from broker hiccups. Agents check for existing output before re-running (resume-from-failure).

🔗 Links

Graph View

Backlinks