Let’s work together.

Whether you need ongoing support or a fresh new direction, I’d love to help you build what’s next.

Get in Touch

Tuinzie

An experimental A.I. concept for the garden industry, evolving from proof of concept to full MVP platform

Tuinzie began as an experiment to explore what modern A.I. can mean for the garden industry. The idea is simple: help people design and visualize their garden instantly. No waiting for expensive 3D renders, no technical knowledge, and no uncertainty. Just real products, realistic designs, and instant inspiration.

The first milestone is the MVP: a tool that allows users to generate credible garden inspiration without uploading their own photo. The focus is on showing real garden materials, starting with tiles, represented with at least 75 percent accuracy. This means correct patterns, surface textures, sizes, and realisti

About the project

The core of Tuinzie is the editor. It lets users explore their future garden with real materials. The editor currently works with manual masking: • upload a garden photo • mask the area by hand • select a product (tile, material, color) • receive four variations At this stage the results land around 50 percent accuracy. In most cases one of the variations matches the chosen product well enough to give a realistic impression. Improving accuracy requires experimenting with conditioning methods, depth workflows, and training the system with more reference images. In the future this will become a guided experience where users receive multiple high-quality concepts in seconds — a task that traditionally costs between €500 and €1500 per render. This experiment shows that 3D-level visualization can be produced on demand and for a fraction of the time and cost.

Tools used

  • Next.js
  • TypeScript
  • Nano Banana
  • ComfyUI
  • Depth Anything
  • Segment Anything
  • LoRAs
  • Language Models
  • Image Models
  • Masking Techniques

Tags

A.I. ExperimentMVPGarden IndustryImage GenerationProduct VisualizationChatbot IntegrationCreative Technology

Auto-Detection Systems

Removing manual masking

To remove the need for manual masking, I have been testing several auto-detection and segmentation systems:

  • SAM2 (Meta) for object and surface segmentation
  • Depth Anything for depth mapping
  • ComfyUI workflows that combine detection + depth + material replacement

The goal is to automatically detect grass, pavement, borders, facades, and other elements so the user only needs to fine-tune the mask. This will significantly reduce friction and create a smoother, more intelligent editor.

This feature will not be included in the MVP but will become a crucial upgrade afterward.

Tuinzie 3
Tuinzie 4

Smart Design Assistant

A.I.-powered guidance

Tuinzie will eventually include a smart garden design assistant that learns the user's style and guides them step by step:

  • selecting materials
  • planning layouts
  • understanding proportions
  • matching colors and textures
  • predicting style preferences

This assistant will run on LLMs combined with RAG, vector search, and a garden-specific knowledge base built with PostgreSQL and Supabase. It will give instant, accurate, and personalized advice — no human designer required.

Vibecoding Architecture

A.I.-assisted development

An important part of this project is the way it is built. I experiment with vibecoding, using Cursor, Sonnet, GPT-5 Codex, and Claude to generate architecture, components, and logic.

But this is not "quick and dirty" coding. I combine A.I. coding with:

  • GitHub version control and branch workflows
  • Local + sandbox + live environments
  • A VPS server for hosting
  • A dedicated A.I. server for heavy tasks like ComfyUI and SAM2
  • Next.js, React, TypeScript, Tailwind CSS
  • Python scripts for A.I. integrations

This combination allows me to build extremely fast while keeping the system stable, secure, and scalable. Every feature is developed, tested, and deployed with proper engineering discipline.

Tuinzie 5
Tuinzie 6

Advanced Workflows

ComfyUI and material accuracy

To improve material accuracy I installed ComfyUI on my server and began testing advanced workflows using:

  • Flux
  • Stable Diffusion
  • Custom checkpoints
  • LoRA training
  • Depth + segmentation pipelines

These workflows help me understand how to preserve texture, alignment, and material realism when generating or editing gardens. This is a long-term effort and a crucial part of reaching 75 percent+ product accuracy.

RAG Knowledge Base

Garden expertise at scale

I have set up a RAG workflow that allows unstructured garden knowledge to be imported into Supabase and used for fast, context-aware answers.

As the dataset grows, the assistant will learn more about:

  • materials
  • garden types
  • tile specifications
  • layout strategies
  • seasons, maintenance, lighting, and plant compatibility

This will allow the chatbot to behave like a real garden expert.

Tuinzie 7
Tuinzie 8

Continuous Experimentation

Discovering new tools

HuggingFace has become my discovery engine for new tools. I test everything from OpenCLIP to segmentation models and large neural networks. Many of these models run on high-performance A.I. infrastructure to allow rapid experimentation.