How This Blog Is Made
An AI-assisted pipeline with a human at the wheel — how 180+ feeds become drafted, illustrated, human-approved posts.
Every post on this blog tagged ai-assisted was produced by an automation pipeline I designed, built, and operate in my home lab — with me making the editorial calls. This page explains exactly how that works, because if I'm going to write about AI systems all day, you deserve to know how much of this site is one.
The short version
Software finds the stories, drafts the posts, and builds the images. I choose what gets covered, review what gets written, and approve every single post before it publishes. Nothing goes live without a human tapping "approve."
The pipeline
1. Signal
A self-hosted service monitors 180+ curated RSS feeds — Microsoft engineering blogs, Azure updates, Fabric release notes, community sources — and assembles a daily digest. From that digest, the system proposes blog topics and checks them against everything already covered so it doesn't repeat itself.
2. Draft
An n8n workflow orchestrates research and writing through Azure OpenAI. The writer works from a structured outline with a defined voice and an assigned angle, then the draft lands in Ghost as — and only as — a draft.
3. Images
The images are the part I'm proudest of. Instead of asking an image model to hallucinate a graphic (and misspell "Kubernetes" in the process), the pipeline composes HTML using a locked design system and a library of real product icons, then renders it to PNG/GIF with a headless browser. A vision model reviews every image against a quality rubric — spelling, layout overlap, whether it actually represents the post — and sends it back for another attempt if it fails. AI never draws a logo or a face here.
4. Gate
When a draft is ready, I get a push notification with a preview. I read it, and I approve it, edit it, or kill it. That's the deal: the machine does the assembly line, the human owns the byline.
Why build this?
Because the pipeline is the content. I work on Azure, AI, and data systems for a living — running my own production-grade agentic system, with real failure modes and real governance decisions, teaches me more than any demo could. When I write about AI governance, quality gates, or human-in-the-loop design, it's because I run those problems in my basement every day.
What's always human
- Topic selection and the editorial calendar
- The final approve/reject on every post and every image
- Opinions, hot takes, and anything from personal experience
- The architecture and every line of the pipeline itself
Questions about the setup? The broader infrastructure it runs on is documented on The Lab page, and I'm always happy to talk shop on LinkedIn.