Building practical AI in public
Experiments, coding agents, AI workflows, honest model comparisons, and systems that actually make it to production — for people building with AI.
§Tools
Small, sharp tools that each do one useful thing — live on their own subdomains, not stuck in a demo folder.
Hook Machine
Built in ~3 days
Blank-page syndrome for anyone who has to write hooks.
Enter your website once. Hook builds a permanent Business Brain, then generates original, business-specific marketing hooks — contrarian angles, story openings, ad headlines, and scroll-stopping scripts — on demand.
Prompt Inspector
Live · actively iterating
You're paying frontier-model prices for prompts a cheaper model could nail.
An OpenAI-compatible proxy that inspects each prompt with a cheap model first, then routes it to the least expensive tier that can actually handle the job — so you stop overpaying for questions that never needed a frontier model.
§The work
Different projects, one pattern: put AI to work on something real, then show how it was done.
Small products that do one useful thing well and ship on their own subdomain.
Putting Claude Code and other agents on real builds to see where they shine and where they break.
Chaining prompts, tools, and models into workflows that actually finish the job.
Wiring AI into the boring, repetitive work so it runs without a human babysitting it.
Honest, build-based comparisons — which model earns its cost on a real task, not a benchmark.
Implementing MCP servers, tool calling, and AI frameworks in production, not slideware.
Idea to shipped in days, documented start to finish — including the parts that go wrong.
Useful software worth sharing, published for anyone to read, run, or steal.
§Build in public
Every project comes with the honest version: what shipped, what fell over at 1am, and what I'd do differently next time.
A marketing-hook generator I built fast and used as a live coding-model comparison. One agent got it most of the way in a day; a second had to finish the job and harden it for production.
An OpenAI-compatible proxy that inspects each prompt with a cheap model first, then routes it to the least expensive tier that can actually handle the job.
The site you're reading was built by a coding agent working from a spec: a full personal-brand site, a design system, SEO, and a Vercel deploy — in one pass.
§What you get
No motivational threads, no reposted AI news. Just the work, taught from actually doing it.
The whole build — the shipped version, the part that broke at 1am, and what I'd do differently. No highlight-reel-only.
How the tools actually get made: coding agents, MCP, tool calling, real workflows. Taught from implementation, not theory.
Which models earn their price, which frameworks are hype, and what I'd actually reach for. Evidence, not vibes.
Small, useful software shipped on subdomains. Some become products; some become cautionary tales worth copying.
§Follow the work
New tools, honest build reports, model comparisons, and workflows worth copying — as they happen. If it's here, it's because it actually got built.