How This Site Works
The Short Version
This site is an AI-assisted publishing system. I work with an AI assistant who manages the entire content pipeline — from research and drafting to formatting, image generation, audio production, and publication. Everything you see here runs largely unsupervised, which means you might occasionally spot an error or something that doesn't quite land. If you do, please leave a comment. My assistant monitors them regularly and we'll review and fix issues as they come up.
Listen to This Article
Choose your preferred voice:
Why Build It This Way?
I'm a Principal Data Engineer in Dublin. My day job involves building data pipelines, managing infrastructure, and helping organizations move from legacy systems to modern cloud platforms. I started this site as a playground — a place to experiment with AI tools in a real-world context, not a theoretical one.
The goal isn't to sell a product or build an audience for monetization. It's to:
- Share what I'm learning about AI infrastructure and data engineering
- Demonstrate AI-first workflows in practice, not just in theory
- Test and benchmark new models as they launch — if a frontier lab drops something interesting next Tuesday, we'll have fun testing it and publishing thoughts by Wednesday
- Build in public — transparency about how the sausage is made
The Technical Stack
Content Management: Ghost
The site runs on Ghost — an open-source publishing platform. I self-host it using Podman. Ghost handles the CMS layer: posts, pages, tags, SEO metadata, and the email newsletter system.
Text-to-Speech: Kokoro Model
Every article includes audio versions in three different voices:
- Sky (
af_sky) — Default American female, crisp and clear - Onyx (
am_onyx) — Deep American male, authoritative tone - Puck (
am_puck) — Playful American male, conversational feel
These are generated using the Kokoro TTS model, which offers 50+ voices across multiple languages and accents. I chose this specific trio to give readers options — some people prefer listening to female voices, others male; some want serious delivery, others something lighter. The Kokoro model is fast, inexpensive, and produces natural-sounding speech that's easy to listen to for technical content.
The audio appears at the top of each article in native Ghost audio cards, so you can listen directly in the browser or download the MP3.
Image Generation: Chroma and HiDream
Hero images and illustrations are generated using AI image models:
- Chroma — Fast, cheap editorial illustrations (~$0.01 per image)
- HiDream — Softer, dreamier compositions when that fits the article
I don't specify the exact provider for these — the model names are what matter. The aesthetic is consistent across the site: warm terracotta and cream palettes, geometric editorial style, clean lines. This visual consistency helps the site feel intentional rather than thrown together.
Research Workflow
Content ideas come from a few places:
- ArXiv papers — I track ML systems, data engineering, and infrastructure papers
- Frontier lab releases — OpenAI, Anthropic, Google DeepMind, etc. When they ship something significant, we test it
- Client work insights — Sanitized learnings from actual data engineering projects
- Tool testing — New databases, orchestration systems, observability platforms
The research process involves my AI assistant pulling sources, summarizing papers, and helping identify angles worth exploring. I review and refine everything before it goes live, but the heavy lifting of synthesis and first drafts happens collaboratively with AI.
Publishing Cadence
New articles publish every Wednesday. This schedule is intentionally modest — weekly is frequent enough to build momentum and demonstrate consistent output, but not so aggressive that quality suffers or it becomes a burden. If there's nothing worth saying one week, we skip it. The goal is signal, not noise.
What My AI Assistant Actually Does
I want to be specific about the division of labor here because "AI-assisted" can mean anything from spell-checking to full automation. In our workflow:
The AI assistant handles:
- First-pass research and source gathering
- Drafting articles from my notes and outlines
- Generating TTS audio in all three voices
- Creating hero images from prompts I provide
- Formatting content for Ghost (HTML, frontmatter, etc.)
- Uploading media and publishing posts
- Monitoring comments and flagging issues for review
I handle:
- Strategic direction (what topics, what angles)
- Final editing and fact-checking
- Technical validation (code, architecture decisions)
- Voice and tone (making sure it sounds like me)
- Responding to comments that need human context
The system runs unsupervised in the sense that I don't micromanage every step. I give direction, the assistant executes, and I review before it goes live. Over time, as the assistant learns my preferences, the review process gets faster and lighter.
If You Spot An Error
Because this is an automated, AI-assisted workflow running largely unsupervised, mistakes will happen. A hallucinated fact. A misinterpreted paper. A sentence that doesn't quite parse.
If you catch something, please leave a comment on the article. My assistant monitors all comments and flags anything that looks like a correction or question. We review these regularly and will fix errors as they're identified.
Consider it collaborative quality control. Your corrections make the content better for everyone who reads it after you.
What's Coming
The current focus is data engineering infrastructure, particularly for healthcare and regulated industries — that's where my recent work has been concentrated. But the beauty of an AI-assisted workflow is we can pivot quickly. If a new model drops that changes how we think about embeddings, or a frontier lab makes a notable mistake worth analyzing, we can turn around an article in days, not weeks.
Subscribe if you want to follow along. Or don't — this site works either way. It's a playground first, a publication second.
