Own AI product (SaaS)
GardenPilot
A single garden photo becomes AI visualisations, a concept and an actionable garden plan – orchestrated by dozens of specialised AI agents.
Visit project30+
AI agents in production
01 · Starting point
Garden planning is expensive and abstract for homeowners: a paper plan is hard to imagine, and a landscape architect is out of reach for many. At the same time, landscaping companies lose a lot of time on initial consultations that never turn into jobs.
02 · Solution
- 1 Users upload a photo of their garden. An orchestrator agent guides them through four phases: vision, concept, planning and implementation.
- 2 Specialised sub-agents produce photorealistic visualisations, planting concepts, a project folder as PDF and optionally a CAD plan.
- 3 Landscaping companies are connected as partners and receive pre-qualified enquiries with a finished concept.
- 4 The image pipeline signs every AI image with C2PA provenance data and labels it as AI-generated.
03 · What I learned
A prompt is not a guardrail
Rules that must never be broken belong in code: at each step an agent only sees the tools it needs, values are enforced in the schema, and text that must stay unchanged is never sent to the model.
Dry run first, real data second
Before a model changes data, it runs against a record that must not change. Only when nothing happens there does it touch production data.
Measure success by the result
A success message is not proof. A deployment only counts as done once the new content has actually been verified live.