Project Overview
Riff AI News is a fully automated video news production pipeline. Six times a day it wakes up, researches the latest AI and technology stories, writes a broadcast-quality script using Claude AI, generates a narration audio track via TTS, sources stock footage from Pexels, assembles everything into a finished MP4 news segment, and produces a separate YouTube Shorts version. No human touches the production process.
Key Features
- Fully automated 6x daily production schedule -- cron-triggered pipeline runs at fixed intervals (midnight, 7am, noon, 3pm, 6pm, 10pm) Melbourne time.
- AI research and script writing -- Claude AI researches current topics via web search and writes structured broadcast scripts with hooks, scene breakdowns, and narration copy.
- Text-to-speech narration -- XTTS and edge-tts generate natural-sounding voiceover audio for each scene.
- Automated stock footage sourcing -- Pexels and Pixabay APIs supply relevant B-roll clips matched to each script scene by keyword.
- Video assembly via ffmpeg -- individual scene clips with audio are concatenated, timed, and rendered into a finished MP4 with scene transitions.
- YouTube Shorts variant -- a separate vertical-format short is auto-generated alongside the main video.
- Crash recovery and orphan watchdog -- a watchdog process checks every 5 minutes for stalled jobs and resumes them automatically from their last checkpoint.
- Nightly quality review -- an automated overnight process reviews recently produced videos and logs quality metrics.
- Weekly channel report -- a weekly summary of production output, view metrics, and topic performance is generated automatically.
Technical Highlights
The pipeline is a Node.js scheduler (scheduler.js) with modular stages: researcher, scripter, asset-fetcher, voice-gen, render-dispatcher. Heavy rendering is offloaded over SSH/SCP to a dedicated GPU workstation. All job state is persisted as JSON, enabling reliable resume-on-crash. The Claude API handles both research (web search mode) and script writing, with a full conversation context passed between stages for coherence.