> pipeline/stages

how it works

Six stages, fully automated. We run them on our infrastructure — you never install anything. Human oversight optional at each.

• 01 topic-research

  Telegram messages, RSS feeds, and Tavily web research ingested and clustered by topic. AI picks the editorially strongest, most diverse cluster.

  telegram apitavilygpt-4o-mini clustering
• 02 script-generation

  A LangGraph critique loop drafts the script, then self-critiques for banned modifiers, journalistic accuracy, and narrative arc. Prompts live in Django, editable via UI.

  langgraphgpt-4.1critique loop (3 iterations)
• 03 visual-selection

  For each segment, 5 image candidates fetched from DuckDuckGo. gpt-4o-mini (vision) picks best fit given full narrative context.

  duckduckgo imagesgpt-4o-mini visiondisk cache
• 04 voice-synthesis

  OpenAI TTS (tts-1-hd, onyx) generates audio per segment. ffmpeg strips leading silence, normalizes to EBU R128 (−16 LUFS).

  openai tts-1-hdffmpeg loudnormsilenceremove
• 05 video-assembly

  Two paths: ffmpeg ken-burns slideshow for landscape long-form, or Remotion React compositions for vertical shorts.

  ffmpegremotionreact compositions
• 06 captions-and-delivery

  Whisper transcribes audio to word-level SRT. Captions burned in via ffmpeg. The full production folder ships to your Google Drive or Dropbox — final landscape + vertical MP4s, full segmented script (JSON), word-level caption timings, per-segment images (with vision-model reasoning), per-segment voiceover audio (mastered + raw), thumbnail PNG, metadata JSON with suggested title / description / tags. You can re-cut, repurpose images on socials, translate captions, or just upload the MP4 — the assets are yours.

  whisper verbose_jsongoogle drive / dropbox syncfull run-folder export