I built an end-to-end content pipeline that generates an SEO-optimised blog post and publishes it to WordPress in under fifteen seconds. The interesting part is not that AI wrote an article. It is the cost and latency profile that made it worth running at all.
The problem with the obvious approach
Content is expensive in two directions. Freelance writers cost per article and take days. Premium model APIs are quick but priced per call, roughly fifteen to fifty cents an article once you account for a decent prompt, and several of them carry monthly minimums. For a business that wants steady output across a handful of sites, neither shape works.
The stack
- n8n for workflow automation, self-hosted in Docker so there is no per-execution billing.
- Groq running Llama 3.3-70B for generation. The free tier allows 14,400 requests a day, which is far more headroom than a content operation needs.
- The WordPress REST API for publishing, so no custom endpoint or plugin is required on the receiving site.
- PostgreSQL for state and record keeping.
What it achieves
- Ten to fifteen seconds per article, against sixty or more through GPT-4.
- Zero AI cost, against fifty to five hundred dollars a month on the alternatives.
- Capacity for up to 14,400 articles a day, which is enough to keep forty blogs running around the clock.
The speed difference matters more than it first appears. At sixty seconds a call you think of generation as a batch job. At twelve seconds it becomes something you can trigger from a webhook and wait on.
The engineering, not the prompt
The prompt was the easy half. What took the time:
- Multiple REST integrations. Groq and WordPress speak different dialects of authentication and error reporting, and both need handling properly.
- Webhook-driven execution so the pipeline reacts to events rather than polling on a timer.
- Error handling and retries. Free tiers rate limit, networks fail, and a pipeline that gives up on the first non-200 is not a pipeline.
- A modular workflow so a step can be swapped without rebuilding the chain.
- Docker networking. Getting containers to reach each other and the outside world reliably was the single most annoying part of the build, and worth mentioning because it always is.
Where it goes next
V2 is in progress: automated image generation and upload, smart tag and category assignment, multi-language output, and scheduling so posts land on a calendar rather than all at once.
Why I am sharing it
Open-source tools give you room to build, experiment and eventually scale something into a product. The whole system is documented and running in production. If you are weighing up AI content workflows for WordPress, the lesson worth taking is that model choice is a cost decision as much as a quality one, and the free tier of a fast open model covers more ground than people expect.