The moat wasn't the model: what building Hadaa taught me about AI defensibility
When we started building Hadaa, the plan sounded almost too simple: take a photo of someone's yard, run it through image-to-image generation, hand back a redesigned yard. In 2026 the generative part is basically a commodity — the renders look photorealistic out of the box.
That turned out to be the easy 20%. Here's what actually made the product defensible, in case it's useful for anyone building in the "AI does X" space.
The demo everyone can build. A raw diffusion model gives you a gorgeous render and nothing else. We could ship that in a weekend, and so could competitors. When we benchmarked 11 tools, almost everyone had stopped there. The problem: those pictures are unbuildable. The model draws what looks lush, not what grows — it'll cheerfully put tropical palms in a render for a yard that freezes solid every winter, and half the design is dead by spring.
Where the actual work was. The defensibility came from the boring layers around the model: (1) a constraint engine over generation — our "Biological Engine" resolves every plant to a real species checked against the user's USDA hardiness zone, rainfall, and frost dates. It's a data-and-rules system, not a prompt, which is exactly why it's been the hardest thing for competitors to copy. (2) Structured output, in reverse — turning an approved render back into a contractor blueprint and a bill of quantities. Pixels to priced data. (3) A pipeline, not a chat — one photo in; 22 renders + planting guide + blueprint + BOQ out, in under a minute.
The lesson I keep repeating to other founders: in most AI products right now, the model is the commodity and the moat is the layer that constrains it to real-world truth and emits something actionable.
What it turned into. On the homeowner side it's pay-per-project ($12, no subscription) — deliberately cheap, because the automation means we're not selling hours. On the pro side it's a closing tool for landscapers: generate the vision live in a client consultation and hand over a white-labeled buildable plan on the spot (full pro breakdown here).
Happy to answer anything about the build — the constraint-engine-over-a-generative-core pattern especially. It generalizes way beyond gardens. It's live at hadaa.app if you want to poke at it.
