What I Built
TrailMate AI is a proposed offline-first nature companion designed to make outdoor walks more engaging, curious, and mindful.
We spend a lot of time looking at screens. What if AI could help us spend less time on them?
TrailMate is designed around a simple idea: use AI to prepare people for an outdoor experience, then let the real world become the main event.
Users will be able to choose a walk duration, receive nature observation missions, discover things to look for around them, and record their observations in a personal nature journal.
Think of challenges like:
- Find three different leaf shapes.
- Spend two minutes listening to the sounds around you.
- Observe how sunlight changes across a tree.
- Notice a small detail in nature that you would normally walk past.
- Write down one thing you learned during your walk.
The goal isn't to turn nature into another screen-based game. It's to make the screen useful for a moment, then encourage people to put their phones away and explore.
Demo
Status: In development.
- Live demo: [Add deployed demo URL]
- Demo video: [Add video URL]
- Screenshots: [Add screenshots after implementation]
I'll update this section with a working demonstration once the prototype is ready.
Code
Repository: [Add your public GitHub repository URL]
The project is intended to be built with a focus on reproducibility, local execution, and clear documentation.
The repository will include setup instructions, the application source code, model configuration, and instructions for running the AI locally.
How I Built It
My goal is to build TrailMate around open-weight AI rather than making a closed, remote AI API the core dependency.
Proposed architecture
Frontend: Next.js, TypeScript, and Tailwind CSS.
AI: A small Qwen instruction-tuned model, selected according to hardware requirements, output quality, and its applicable license.
Local inference: llama.cpp, using a compatible quantized model.
Nature knowledge: A small, curated local dataset containing observation ideas and basic educational information.
Persistence: Local browser storage for walk history and journal entries.
The application will combine predefined observation missions with locally generated prompts. The curated content will provide a predictable fallback if the model is unavailable or produces an unsuitable response.
Rather than asking a model to invent every fact about nature, the application will use a limited knowledge source and encourage users to verify uncertain observations.
An important technical goal is to keep inference on the user's device. Once the application, model, and required data are available locally, the core experience should not need an internet connection.
I'll document the actual model, quantization settings, hardware, and offline test results after implementation.
Why Does Open Innovation Matter?
For TrailMate, open innovation isn't just about choosing a different AI model. It changes how the product can be built and used.
1. Privacy by design
Outdoor journals can contain personal reflections and information about where someone spends their time. A local-first design can keep these records on the user's device instead of requiring a remote AI service to process every interaction.
2. Less dependence on closed APIs
A remote AI API can introduce usage limits, network dependencies, and ongoing costs. Local inference can reduce those dependencies, although it still requires compatible hardware, storage, and electricity.
3. Freedom to experiment
Open-weight models and open-source inference tools give developers more control over model selection, quantization, prompts, and execution.
That makes it easier to experiment with smaller models, compare their output quality, and adapt the application to different devices.
4. A more accessible approach to AI
A nature companion shouldn't require a constant internet connection just to generate a simple observation prompt. Local inference makes offline use a realistic design goal.
The broader lesson is that AI doesn't always need to be a cloud service. Sometimes, the most useful AI is the kind that quietly supports an activity happening somewhere other than a screen.
My Agent Session
[Add a DevRelay agent session link or embed after saving the actual session.]
This section is optional and will be completed if I capture a suitable development session.
Prize Categories
[Add applicable partner prize categories after checking the official challenge requirements, or remove this section if none apply.]
I'm building TrailMate AI as a solo project for the Touch Grass challenge.
The measure of success won't just be whether the model generates interesting text. It will be whether the application helps someone notice more of the world around them.
Build with AI. Step outside. Touch grass. 🌱
Top comments (0)