Skip to content

← back

Amuma Garden Designer

Explainable layout planning for beginner vegetable and herb gardens

GitHub → Watch the full demo → Read on Medium →

What it does

  • Takes bed dimensions, location (USDA hardiness zone), sun exposure, and crop choices and generates a plantable grid layout instantly — no account required.
  • A deterministic solver scores every candidate layout against spacing requirements, mature plant spread, companion planting conflicts, water needs, and harvest-aisle access. Hard constraint violations are blocked before the AI ever sees the layout.
  • An AI reviewer writes a bounded final narrative — summary, strengths, concerns, and suggested fixes — using only the structured facts the solver already validated. It cannot invent plants, override spacing rules, or add crops you did not select.
  • Outputs a planting table with start method (direct sow vs. transplant), spacing, and timing for each crop, plus a full audit trail of what the app knew, where the data came from, and how it scored each constraint.

Why I built it

Every gardening app I tried either required expert knowledge to fill out the intake form — hardiness zones, root depth, companion planting conflicts — or handed back vague AI suggestions with no way to verify them. The design constraint that shaped Amuma: AI can explain the layout but it cannot determine the layout. Every bed dimension, planting window, spacing rule, and companion conflict has to live somewhere sturdier than a prompt. That discipline turned out to be the hardest part of building a tool a beginner can actually trust.

How it works

User Inputbed size · location · crops · sunPlant CatalogVerdantly plant APInormalized to consistent shapesaudit trail per plant record31 supported cropsDeterministic Solverspacing · mature spreadcompanion conflicts · water needsharvest-aisle accessgenerates + scores candidate layoutsAI Reviewer (bounded)explains tradeoffs in plain languagecannot invent plants · override spacingor add crops you did not selectScored Gridplanting layout with constraintscores and full audit trailPlanting Tablestart method · spacing · timingper crop in the bed
The plant catalog and deterministic solver run independently, then merge. The AI reviewer only sees validated facts — it explains the layout it cannot change.

What was hard

  • The first version asked too much of the user. A beginner does not know their hardiness zone, the difference between direct sow and transplant crops, or what companion planting means. The final intake only asks what materially changes the plan.
  • Drawing the line between solver and AI. Anything with a deterministic right answer — spacing, mature spread, hard companion conflicts, planting windows — belongs in the solver. Anything that benefits from natural language — explaining why two crops conflict, surfacing height and shade tradeoffs — goes to the LLM. Mixing the two in one step erases the audit trail.
  • Evals as internal controls. Three layers run before any layout reaches a user: plant record normalization validation, verification of all supported plants, and crop-mix decision validation. They exist to catch confidence-building errors quietly — the failure mode that only shows up after the demo.

Outcome

  • A scored planting grid for any vegetable or herb bed, generated from bed size, location, sun, and crop choices.
  • A bounded AI narrative that explains the layout's strengths, concerns, and suggested fixes — with no ability to override the solver's decisions.
  • A full audit trail: what the app knew about each plant, where the data came from, and how every constraint was scored.
  • A planting table with start method, spacing, and timing per crop. Live at www.amuma.ai.

Tech

Next.js · TypeScript · Pixi.js · OpenRouter (Qwen 3 32B) · Verdantly plant API · deterministic solver