This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
What if AI became the reason you put your phone away instead of another reason to keep looking at it?
That's the idea behind FIELDCRAFT ðŸ§, an audio-first outdoor mystery game that turns your neighborhood into a real-world adventure.
Most AI applications are designed to keep us interacting with a screen. I wanted to build something different: an experience where AI creates the story, the real world becomes the game board, and physical movement determines what happens next.
In FIELDCRAFT, you're not just reading a story about an adventure. You're walking through one.
Here's how it works:
Choose your adventure. Pick a walking duration and a narrative style, such as Mystery & Noir, Wonder & Nature, or Playful Detective.
Generate a real-world mystery. FIELDCRAFT discovers nearby points of interest using OpenStreetMap and uses Gemma to create a fictional case with an opening narrative, clues, hints, and a finale.
Put your phone away. Clues can be narrated using browser-based speech synthesis, allowing you to listen and explore instead of continuously watching your screen.
Unlock the story by moving. In live mode, the application uses GPS geofencing, location accuracy checks, and repeated qualifying readings to determine whether you've reached the next checkpoint. You can't simply press a button to skip ahead.
Solve the mystery. Reach the final checkpoint, discover the ending, and keep a record of your adventure in the field journal.
FIELDCRAFT also includes a clearly labelled simulated demo mode, so anyone can experience the adventure without physically visiting every checkpoint.
The idea behind FIELDCRAFT
The central design principle is simple:
Gemma writes the mystery. OpenStreetMap grounds it in real places. Deterministic code controls progression. Your feet unlock the story.
The AI generates the fictional narrative, but it doesn't decide whether you've reached a checkpoint. That decision belongs to the application logic.
This separation makes the experience more than an AI-generated scavenger hunt. It combines creative storytelling, open geographic data, audio interaction, and physical-world constraints into a single experience.
FIELDCRAFT is for curious explorers, students, walkers, and anyone who wants to experience familiar surroundings differently.
FIELDCRAFT in Action
Here are screenshots of the actual application running, showing how the adventure looks and feels.
Here:Here:
The walkthrough demonstrates how a player creates an adventure, receives a generated mystery, listens to clues, progresses through checkpoints, and reaches the final reveal.
The application distinguishes between real GPS-based progression and simulated demo arrivals.
Code
GitHub repository: farhan4783/Fieldcraft
The repository contains the frontend, Express backend, AI generation services, geographic utilities, checkpoint-validation logic, fallback narrative engine, tests, and implementation documentation.
FIELDCRAFT is released under the MIT License.
Contributions, experiments with alternative open-weight models, and improvements to the storytelling and exploration experience are welcome.
How I Built It
I built FIELDCRAFT using HTML, CSS, vanilla JavaScript, Node.js, and Express. The project combines open-weight AI with open geographic data and standard browser APIs.
The AI: Gemma 3
FIELDCRAFT uses Google's open-weight Gemma 3 model to generate structured fictional mysteries grounded in nearby points of interest.
I designed the application to support multiple inference paths:
- Local inference: Ollama runs Gemma 3 locally, allowing the story-generation process to run on the user's machine.
- Alternative inference: An optional OpenRouter integration can serve an open-weight Gemma model when local inference isn't practical.
- Deterministic fallback: A built-in procedural narrative engine keeps the experience playable when model inference is unavailable or generated output fails validation.
The AI-generated story follows a defined structure rather than being treated as unrestricted text. The application validates the output and its checkpoint references, attempts to repair invalid output, and falls back to the procedural engine when necessary.
The real world: OpenStreetMap
FIELDCRAFT uses OpenStreetMap's Overpass API to discover nearby points of interest, with Nominatim supporting geographic search.
This gives the story a real-world setting rather than generating an entirely imaginary landscape.
The places provide the setting for the fictional mystery; the application does not claim that its invented story or clues represent verified local history.
The game engine: deterministic code
One of the most important engineering decisions was not allowing the language model to control the game's rules.
The application handles location permissions, GPS readings, checkpoint proximity, validation, and story progression independently of the AI.
In live mode, a checkpoint is unlocked only when the location logic accepts the player's arrival. In simulated demo mode, arrivals are explicitly labelled as simulated.
This makes the creative layer flexible while keeping physical progression predictable and testable.
The experience: the browser
The frontend uses browser geolocation for live movement, Speech Synthesis for narration, Web Audio for interface sounds, and LocalStorage for the field journal.
The application is designed to minimize screen interaction during a walk. However, live place discovery requires network access, and browser narration and GPS behavior can vary across devices.
Why Does Open Innovation Matter?
Open innovation matters to FIELDCRAFT because the model is not simply an external service attached to an otherwise conventional application. The ability to run and experiment with an open-weight model changes how the experience can be built and controlled.
1. AI that can run locally
In local Ollama mode, Gemma runs on the user's machine rather than requiring a closed commercial AI API for story generation.
This provides greater control over the inference environment and avoids sending the story-generation request to a proprietary model provider.
There is an important distinction: live location discovery still requires network requests to geographic services. FIELDCRAFT is not entirely offline in its live mode, and I don't claim that all location-related data remains on the device.
2. Freedom to experiment
Because the model, prompts, output validation, narrative fallback, and game mechanics are separate components, developers can experiment with alternative open-weight models and storytelling approaches without rebuilding the entire application.
The rules governing checkpoint progression remain under the application's control, independently of which compatible model generates the story.
This is one of the strengths of combining open models with conventional software engineering: the model can be creative without being responsible for every decision.
3. Open geographic data and open AI
OpenStreetMap provides the real-world context, while Gemma turns that context into a fictional narrative.
Together, they make it possible to create location-aware experiences without relying on a proprietary AI API for the creative layer.
The broader lesson I took from this project is that open innovation isn't just about access to a model. It is also about being able to inspect, adapt, and replace the components that make an experience work.
4. AI with a reason to disappear
The most important part of FIELDCRAFT isn't the amount of time someone spends interacting with its AI.
It's what happens after the AI generates the next clue.
The screen becomes secondary. The neighborhood becomes the setting. The player has to move through the real world to continue the experience.
I wanted to build an AI that doesn't compete for your attention, but gives you a reason to direct that attention somewhere else.
My Agent Session
The project source code and implementation details are available in the GitHub repository.
I haven't included a public DevRelay agent-session link in this submission.
Prize Categories
- Overall Hacktoberfest Open-Source AI Challenge Week 1: FIELDCRAFT is built around open-weight AI and is designed to get people off the screen and into the real world.
- Best Use of Gemma: FIELDCRAFT uses Gemma as its narrative-generation model, with support for local inference through Ollama.
Top comments (0)