DEV Community

Олександр
Олександр

Posted on

I built an LLM router that picks a different model per article type

I've been quietly running a small network of niche content sites for a few weeks. Two are live on real domains:

  • afixu.com - home improvement and DIY tool guides
  • aceju.com - weekly AI tool awards, editorial format scored across 4 criteria

The niche isn't the interesting part. The pipeline is.

The LLM router

Instead of one model for everything, a router picks per article type:

  • premium/commercial pieces go to a stronger model
  • high-volume informational drafts go to a cheaper, faster one
  • if a provider is rate-limited, it falls through a cascade so generation never stalls

The QA loop

Every draft, before it ships, passes through:

  1. an anti-AI-tells cleanup (80+ regex replacements - "delve", "leverage", "robust", etc.)
  2. a rankability judge scoring depth, specificity, intent-match and trust on an 8-point scale
  3. a structure/schema check (FAQ block, JSON-LD, internal links)

A daily job re-scores everything published in the last 24h and auto-regenerates the failures. The whole thing is Next.js 16 + MDX, one git repo per site.

The honest target: a $30k/mo run-rate in ~24 months across 30 mature sites, monetized through guest-link sales, not ads. Slow on purpose.

Happy to go deeper on the router logic or the QA scoring if anyone's interested.

Top comments (0)