YT Machine

Features

• RSS collection with static and JavaScript-rendered article extraction
• Topic scoring, category rotation, and structured analysis
• Two-story script synthesis followed by deterministic enforcement
• Local pixel-art generation with a managed model lifecycle
• Voiceover, karaoke subtitles, and FFmpeg video assembly
• Idempotent YouTube and TikTok publishing with notifications

Final goal

Take a news feed all the way to a finished short-form video using local models, with each stage checkpointed, inspectable, and independently replaceable.

Engineering

The pipeline separates collection, research, script enforcement, visual generation, audio, assembly, and publishing. A phase-aware runtime keeps one heavy model resident at a time, while JSON artifacts remain authoritative when optional services such as PostgreSQL are unavailable.

Pipeline

1. 01

   Collect RSS feeds and article text, then score topics and rotate categories

2. 02

   Research selected stories and synthesize two structured stories with four beats each

3. 03

   Enforce script structure and build visual prompts before the memory barrier

4. 04

   Generate pixel-art scenes, create voiceover and subtitles, and assemble the vertical MP4

5. 05

   Write manifests and metadata, then publish through the idempotent platform ledger