Micro Haven

Inspiration

We kept coming back to a simple question: what can someone do for nature if they do not have a yard?

A balcony, patio, or windowsill may seem too small to matter, but those spaces are still part of the environment around us. We wanted to make taking a small environmental action feel practical instead of complicated.

That led us to MicroHaven: a way to turn a small everyday space into a tiny native habitat.

What it does

MicroHaven asks a few simple questions about your space, such as its size, sunlight, moisture, watering ability, experience level, and budget.

It then looks at a curated set of native plants and finds a small combination that fits those constraints. Instead of simply listing plants, it tries to make them work together by considering things like bloom timing, ecological roles, space, and growing conditions.

The result is a simple habitat plan showing what to plant, where to place it, what it may support, and how to care for it.

How we built it

We built MicroHaven as a lightweight web application with a local plant-data layer and a deterministic matching engine.

The matching process first removes plants that clearly do not fit the user's conditions. The remaining plants are scored, and combinations of two or three plants are evaluated for compatibility.

We also added an optional LLM layer, but it is not responsible for deciding which plants are suitable. It takes the selected results and turns the underlying information into a more understandable explanation. A deterministic fallback keeps the core product working without the LLM.

We intentionally kept the architecture simple so we could spend our limited time on the actual user experience and matching logic.

Challenges we ran into

One of the hardest parts was avoiding the temptation to make the project “just another AI recommendation tool.”

An LLM can easily produce a convincing plant recommendation, but convincing text is not the same thing as a reliable recommendation. We therefore had to separate factual plant data, matching logic, and AI-generated explanation.

We also had to keep the scope under control. There are many things we could have built — maps, image analysis, accounts, mobile apps, AR, reminders, and more — but adding them would have taken attention away from the core idea.

Accomplishments that we're proud of

We are proud that the main idea works as an end-to-end product rather than just a concept.

A user can describe a real space, the system can reason through the available plants, and it can produce a concrete habitat plan that someone could actually act on.

We are also proud of keeping the recommendation process inspectable. We can see which plants were rejected, how candidates were scored, and why a particular combination was selected.

What we learned

We learned that building an environmental product is not only about having a good environmental idea. The difficult part is turning that idea into something specific enough that a person can actually use it.

We also learned that AI is most useful here when it supports the product instead of pretending to be the product. The recommendation logic and underlying data still need to make sense without an LLM.

Most importantly, we learned that reducing the scope can make an idea stronger. We did not try to solve gardening everywhere. We focused on one small problem and tried to make that experience work well.

What's next for MicroHaven

The prototype currently focuses on small spaces in the Willamette Valley.

The next step would be expanding the underlying plant data and supporting more regions while keeping the same principles: local relevance, transparent matching, small actionable habitats, and recommendations that people can understand.

Eventually, we would like MicroHaven to make it easier for many small spaces to become small pieces of a larger connected habitat.

Built With

• css
• ecological
• environmental-technology
• fastapi
• html
• javascript
• json
• native-plants
• openai
• python
• responsive-design
• rest-api
• vercel