Tiller: an AI lawn coach, built and shipped by one person

Lawn advice on the internet is generic, seasonal-blind, and wrong for half the country. Tiller reads a specific yard in a specific place at a specific moment, then says the one right thing to do next. Here is how it's built, and why that matters if you're thinking about AI in your own business.

The problem

Every lawn question has a "it depends" answer — on your grass type, your state, your soil temperature, and the week of the year. Generic advice ignores all four. The bags at the hardware store are marketing; the forums are folklore; the apps are content farms with a search box.

What I built

Tiller's core is a deliberately unfashionable architecture: the AI is not allowed to decide anything. A rules engine built on university extension agronomy (soil-temperature windows, a grass-by-state validity matrix, product conflict checks) decides what is true. The AI writes the coaching voice on top: plain, warm, specific. If a combination is invalid (St. Augustine in a northern state, seed over a pre-emergent), the engine refuses before the AI ever speaks.

No advice is better than wrong advice — so the engine is built to say no first, and only then say it nicely.

• Recommendation engine — rules gate + LLM voice; the recommendation is the product's primary object, not a feed of content.
• A lawn's memory — every action logged builds a history the coach reasons over; the product's value compounds with use.
• Full commercial rail — Stripe products, prices, webhooks, and edge functions, live and verified in production.

The AI content operation

Tiller's growth surface is a set of state-by-grass lawn guides: 125 pages live in production, every one generated by an AI pipeline, and every one gated. The pipeline transcribes an agronomist-ruled validity matrix (which grasses genuinely belong in which states, with a hard denylist for the traps), a coach-voice evaluation gate that rejects off-brand output, and production checks after deploy. That's the honest version of "AI content at scale": generation is cheap; verification is the product.

What one person shipped

Product strategy, design system, front end, recommendation engine, billing integration, SEO architecture, and the content pipeline: one person, nights and weekends, using the same AI-assisted workflow I now build for clients. That's the actual point of this page. The leverage is real, I use it daily, and it's transferable.