Case study · ShyftLabs · 2025
AI-SEO pipeline for 300+ healthcare sites
An 18-step content workflow that generates meta data, page copy, FAQs and JSON-LD while protecting quality and LLM cost.
The approach
Each site needed useful, on-brand content at a scale that a manual workflow could not sustain. I designed a DAG that produces 15 content keys per page and routes failures or low-quality output through alternate models.
Quality and cost controls
- Automated CORE-EEAT quality judging determines whether content can be approved without an editor.
- 24-hour Redis caching and Bedrock prompt caching reduce repeated requests.
- Atomic tenant and global cost guards prevent one workload from affecting the wider platform.
300+
sites in production
~72%
auto-approval rate
~65%
fewer full-price calls
Frequently asked questions
Why use multi-model fallback?
Fallback keeps the multi-step pipeline progressing when a primary model errors, times out or misses the quality bar.
What does the pipeline create?
It generates page-level metadata, hero copy, FAQs and structured data, among other content keys.