GroundSignal: A Mission, Not a Feed | Local AI That Sends You Outside 🌿

Jesse — Mon, 12 Oct 2026 04:29:51 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass.

Most apps are designed around one question: How long can we keep you here? GroundSignal starts with another: How quickly can we help you leave?

We spend so much time looking at screens that even the tools meant to improve our lives can become another reason to stay glued to them. I wanted to build something different: an AI application that gives you a genuinely useful reason to put your phone down, walk outside, notice your surroundings, and experience something real.

The result is GroundSignal, an open-source, local-first AI prototype whose goal is to make the screen the shortest part of the experience.

What I Built

GroundSignal turns a person's interests, available time, energy level, and accessibility preferences into a four-step outdoor micro-adventure.

Instead of scrolling through recommendations, you tell it what sounds good. It interprets your words using a locally trained machine-learning classifier and produces a field mission designed to be completed away from your screen.

The prototype currently supports five mission families:

The Three-Bird Quest: A twenty-minute birdwatching adventure involving listening, observing, and noticing three birds, calls, or movements without disturbing wildlife.
The Five-Foot Garden: A gentle gardening activity that helps someone care for a planter or small outdoor patch, including stationary-friendly options.
The One-Block Reset: A simple ten-minute neighborhood walk focused on sounds, colors, and awareness rather than distance or performance.
The Curiosity Trail: A longer exploration mission encouraging someone to discover three details they have never noticed before.
The Sky Atlas: A quiet skywatching experience built around noticing light, clouds, shape, and sound, with seated options.

The user chooses a ten-, twenty-, or forty-five-minute experience and an energy level. GroundSignal then adapts the mission steps and their time allocations.

Not every outdoor adventure needs to be a hike. Someone with limited mobility, low energy, or only a few spare minutes should still be able to participate.

There's no account required, no leaderboard, no engagement streak, and no incentive to keep scrolling. Generate a field card, take it with you, and go experience something.

A mission. Not a feed.

The goal of reducing screen time is a design intention, not a measured user outcome. The prototype has not yet undergone real-world behavioral trials.

Demo

Watch the complete five-mission competition walkthrough:

https://drive.google.com/file/d/14nJPXgusNeZjcyHqg5f_CmITfCyTOJ7A/view

The 109-second video demonstrates five distinct requests processed by the actual GroundSignal application, including input choices, model-selected missions, and all four steps of each activity.

Visual storyboard:

https://drive.google.com/file/d/1QVk3co9swtBzwG_RzNt7W4anlAak-pBD/view

The demonstration uses real local model outputs and authentic application captures, not fabricated AI responses or footage presented as a completed outdoor field trial.

Code

Public source-code folder:

https://drive.google.com/drive/folders/1UficDai0h_4p8mZUXB9zjeNaFFBilff-

Downloadable source archive:

https://drive.google.com/file/d/1H0T77jsak6GrSB7FXxnTL75FG14G_pqp/view

The release contains the Python application, model-training examples, inference engine, responsive interface, tests, dependency list, README, and MIT license.

The code is publicly accessible through Google Drive as an interim release, not a version-controlled Git repository. Our GitHub publishing account was suspended, so we chose to make the inspectable source available rather than misrepresent an unavailable repository.

To run the Python version locally, install the dependencies listed in requirements.txt, execute the automated tests with pytest, and start the FastAPI application through Uvicorn on localhost port 8000.

GroundSignal also has a standalone browser prototype containing an exported version of the trained model, intended to run without a separate Python service. The downloadable HTML demonstration is available in the public release folder.

There is no production-hosted application or cloud inference service in this challenge release.

How I Built It

GroundSignal uses real local machine learning, but deliberately keeps the model understandable.

I developed a training set of sixty short, author-written examples across five outdoor-intent categories. A scikit-learn pipeline uses character-level TF-IDF features and logistic regression to classify the user's free-form request.

The classifier selects a mission family. GroundSignal then builds a structured field card using a curated mission template and adapts the steps to the selected time, energy, and stationary preferences.

This is actual model inference rather than a set of keyword-based if/else branches. It is also not generative AI: the model classifies intent, and a transparent template system constructs the experience. That distinction is intentional.

The application uses:

Python and scikit-learn for local classification and inference.
FastAPI and Uvicorn for the localhost application.
HTML, CSS, and JavaScript for the responsive interface.
pytest and HTTPX for automated verification.

Our reconstructed release passed 17 automated tests plus four independent local smoke checks. We also compared browser-local classifier behavior against the Python implementation. These tests support the demonstrated behavior, not claims of broad classification accuracy or real-world impact.

GroundSignal asks users to acknowledge that they've checked local conditions themselves. It does not know the user's weather, smoke exposure, location, trail conditions, or whether an activity is safe. Its model weights are not calibrated safety probabilities.

A future public deployment would require additional privacy, accessibility, safety, and operational review.

Why Does Open Innovation Matter?

The philosophy behind GroundSignal is simple: useful technology shouldn't always demand more of our attention.

Building this around an opaque, metered cloud inference service would introduce unnecessary cost and dependence for a task that can be solved with a small local model.

Open-source tools make it possible to inspect the classifier, review and expand its examples, change the mission library, improve accessibility, and reproduce the behavior without paying for external AI calls.

Developers can add new outdoor activities. Educators can adapt the field missions for learning. Environmental organizations can explore community-science applications. Accessibility advocates can help make the system useful for more people.

And because the intelligence can operate locally, the design does not inherently depend on sending someone's personal interests to a cloud AI provider.

That is the part I find most exciting. The objective isn't to maximize the number of AI conversations. It's to help a person do something meaningful and then step away.

What Comes Next

The current prototype is intentionally bounded to five curated mission families. There is plenty left to investigate.

We want to evaluate whether the field cards genuinely help people reduce screen time, expand the accessibility options, add more varied activities, and test how well the small classifier handles unfamiliar language.

Future possibilities include responsibly sourced weather intelligence, wildlife observation, environmental education, and community-science missions. Those integrations are future development directions, not features of this submission.

The first goal was to make the premise real: type a sentence, receive an outdoor mission, and put the phone away.

We built that working foundation.

Prize Categories

Overall Hacktoberfest Week 1 Challenge only. No undisclosed partner technology or partner-prize eligibility is claimed.

AI-Assistance Disclosure

GroundSignal was developed with substantial AI assistance for coding, design, drafting, and quality assurance, under human project direction. The documented tests refer to the local prototype. We do not claim production deployment or verified outdoor user outcomes.

THE WORLD IS STILL OUT THERE.

We built AI to help you find it. 🌿

DEV Community: Jesse