Screenshot coming soon
Reno Studio — making an AI image tool usable by homeowners
- Problem
- The underlying model can produce a photorealistic renovation preview, but only if the prompt is precise. Homeowners don't write precise prompts — they say “make my driveway nicer.” The gap between what the user can express and what the model needs was the whole product problem.
- Approach
- Replaced the free-text prompt box with a guided flow that walks the user through real supplier catalogs — stone, then colour, then laying pattern — and composes the technical prompt from those picks behind the scenes. Users are choosing products they can actually buy, not describing an aesthetic. Generalized it to handle two suppliers with different catalog structures, and built the step count to be dynamic so new renovation categories can add their own steps.
- Decision
- Serve catalog imagery from object storage rather than the repo. More than a hundred swatch, hero, and laying-pattern images (about 17 MB) were about to go into the shared git history — moving them to R2 and stripping them out before the first push kept the repo to code, not binaries.
- Outcome
- Guided studio flow shipped with test coverage on the material picker interactions. I also wrote the design specs and implementation plans the work was built from.
- Stack
- Next.js
- TypeScript
- Google Gemini
- Cloudflare R2
- Vitest
- Biome