This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
Garden Assistant — Local AI for Real-World Gardening
Garden Assistant is an offline-first gardening assistant that helps people decide what to plant, when to plant it, and what to do in their gardens throughout the year.
The idea is simple: instead of spending time searching for gardening advice online, get a practical plan and go outside to put it into action.
The application provides planting recommendations based on regional frost dates and generates a simple field checklist with planting tasks, soil checks, and observations to make outdoors. One of its tasks is to find a pollinator on a flower or a pest on a leaf.
The project separates factual calculations from AI-generated explanations. Python handles the gardening calculations, while the AI turns those facts into short, understandable advice.
Demo
Example: AI-generated gardening advice
Running garden.py iowa produces:
Plant garlic cloves 6 inches apart for next summer. Clear spent plants and roots out of the beds and add a layer of compost or leaves over the empty beds. Order seeds for next spring while varieties are in stock.
Example: Outdoor field checklist
Running garden.py iowa --checklist produces:
Garden checklist - iowa - 2026-10-11
PLANT (3)
- [ ] garlic
- [ ] cover crop
- [ ] compost for beds
CHECK (1)
- [ ] soil: dig 5cm down and squeeze -- is it crumbly, not smeared?
LOOK (1)
- [ ] find one pollinator on a flower, or one pest on a leaf
Nothing outdoors until 2027-05-05.
The checklist is generated directly from Python's calculations and works without loading the AI model.
Watch the demo: Garden Assistant — Video Demo
Code
Source repository: Garden Assistant on GitHub
The project includes frost-date calculations, planting recommendations, local AI inference, and automated tests for regional data, date calculations, overrides, and model-generated advice.
How I Built It
The project uses Qwen3-1.7B-Q8_0, an open-weight model distributed under the Apache-2.0 license, through llama-cpp-python.
The model runs locally from a GGUF file. No paid AI API or remote inference service is required.
The architecture has three main parts:
-
Python (
frost.py): Calculates frost dates, determines planting windows, and generates the outdoor checklist. -
Qwen3 (
garden.py): Uses the calculated facts to generate short, practical gardening advice. - Automated tests: Check the calculations and verify that generated advice remains grounded in the supplied facts.
A key design decision was to keep the model away from date arithmetic. A small language model can make mistakes with calculations, so Python determines the factual results before the model receives them.
I tested the application on a machine with an NVIDIA RTX 2050 (4 GB VRAM) and 16 GB RAM. My current llama-cpp-python build is CPU-only, so inference runs on the CPU rather than the GPU. Generating an answer takes approximately 10 seconds on my 12-core CPU.
The application runs locally, and the gardening checklist does not require model inference.
Why Does Open Innovation Matter?
Open innovation makes this kind of local AI application possible without depending on a proprietary inference API.
With an open-weight model, developers can run inference on their own machines, inspect the implementation, change models, and adapt the application to their needs. Gardening queries can be processed locally without being sent to a remote AI inference provider.
It also allows each component to do what it does best. Python handles the calculations that need predictable results, while the language model makes those results easier to understand.
For a small project like this, local inference provides independence from paid API services and remote inference availability. It also makes experimentation accessible to developers working with modest hardware.
The goal is not to put AI everywhere. It is to use AI where it helps, then get people away from their screens and into the real world.
My Agent Session
No DevRelay agent session is included in this submission.
Prize Categories
No partner-specific prize category is being entered.
Limitations
- The project currently supports 40 US regions, not worldwide gardening locations.
- Frost dates are regional averages, not live weather forecasts. Actual conditions can vary by year and location.
- AI-generated advice depends on the quality of the supplied facts and the model's ability to follow instructions.
- The current inference configuration uses the CPU; GPU acceleration is not enabled.
Garden Assistant is a small experiment in using open-source AI to make real-world activities easier rather than creating another reason to stay online.
The screen helps prepare the plan. The garden is where the work happens.
Top comments (0)