DEV Community

Ayush Dinda
Ayush Dinda

Posted on

TrailBuddy

#devchallenge #hf26challenge

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

What I Built

TrailBuddy is an open-source, offline-first outdoor companion engineered around a single radical product principle:

"Use your screen to plan. Then put the screen away and go outside."

Most modern AI apps are built as conversational time-sinks—they reward endless prompt loops, infinite chat threads, and digital addiction. TrailBuddy does the exact opposite. It treats the screen as a temporary launching pad:

  1. Personalized Outdoor Mission Planning: In under 30 seconds, users configure their available time (15m to 2+ hours), outdoor activity (walking, hiking, bird watching, photography walk, mindful outdoor time), mood, difficulty level, and interests (flora, architecture, wildlife, mindfulness).
  2. Local AI Mission Synthesis: Instead of generic motivational essays, a local open-weight model (Qwen3 or Llama 3.1 via Ollama) generates structured, concrete observation quests with sequential challenges (e.g. noticing bark textures, mapping ambient sounds, or tracking cloud paths) and a strict screen-free goal (e.g., 40 minutes of uninterrupted pocket time).
  3. Screen-Free by Design: The moment the user clicks "Start Mission", the app collapses into an intentionally minimalist, high-contrast countdown screen with a prominent visual reminder: "Your screen time is over. Go outside. We'll be here when you return."
  4. Mindful Completion & Offline Reflection: Upon returning, users record how the outing felt and journal what they physically noticed. The record is permanently archived in a local SQLite database without any cloud surveillance.

Who is it for?
Developers, students, remote workers, and outdoor enthusiasts experiencing screen fatigue, digital burnout, and cognitive overload who want a gentle, structured nudge to reconnect with the physical world.

Demo

  • Live Local Demo URL: http://localhost:5173
  • Backend API & Swagger Docs: http://localhost:8000/docs

Key User Experience Walkthrough:

  • Hero & Landing: Nature-inspired design language (moss, earth, sky tones) with pure CSS vector art, zero external asset dependencies, and an instant local AI health indicator.
  • Multi-Step Questionnaire: Accessible 6-step form with visual progress indicators, keyboard accessibility, and cycling AI planning animations.
  • Mission Preview: Outlines mission objectives, sequential sensory challenges, screen-free time targets, and situational safety guardrails.
  • Active Countdown Mode: Intentionally distraction-free screen with pause/resume and full localStorage persistence across browser refreshes or device sleep.
  • Reflection Archive: A private SQLite history journal capturing feelings (Energizing, Relaxing, Challenging) and sensory observations.

Code

GitHub logo AyushDinda / TrailBuddy

TrailBuddy — Your Offline Outdoor Companion 🌿

"Use your screen to plan. Then put it away and go outside."

TrailBuddy is an open-source, AI-powered outdoor activity companion built specifically for the Hacktoberfest Open-Source AI Challenge Week 1: "Touch Grass".

Instead of trapping you in an endless chat loop or addictive digital feed, TrailBuddy harnesses the power of local open-weight AI (Qwen3 via Ollama) to generate realistic, mindful, screen-free outdoor adventures tailored to your available time, mood, and surroundings. Once your adventure begins, the application intentionally becomes minimal—providing only a quiet countdown timer and observation prompts—actively encouraging you to tuck your phone in your pocket and step into the physical world.

Why TrailBuddy? (The "Touch Grass" Problem)

Modern consumer technology is engineered around digital retention: infinite scrolling feeds, notifications, and artificial chat personalities designed to keep eyes glued to glass screens.

The Hacktoberfest Week 1 challenge posed an…

View on GitHub

Repository Tech Stack:

  • Frontend: React 18, TypeScript, Vite, Tailwind CSS, Lucide React, React Router.
  • Backend: Python 3.13, FastAPI, Pydantic v2, SQLAlchemy, Uvicorn, HTTPX.
  • Database: SQLite (trailbuddy.db).
  • AI Engine: Ollama running open-weight Qwen3 / Llama 3.1 locally on localhost:11434.
  • Testing: pytest, FastAPI TestClient, and end-to-end automated integration tests.
  • License: MIT License.

How I Built It

TrailBuddy was architected from day one to keep the AI meaningful, private, and strictly offline:

1. Local Open-Weight Inference with Ollama & Qwen3

Instead of calling a proprietary cloud API over the internet, TrailBuddy sends structured prompts directly to a locally hosted Qwen3 (or Llama 3.1) model running on Ollama (http://localhost:11434). The backend enforces strict JSON schema formatting directly in the model's generation parameters, ensuring zero conversational preamble and instant client-side parsing.

2. Guardrailed Mission Engineering

The AI system prompt enforces strict real-world safety heuristics:

  • Zero Trespassing: Never instructs users to cross private land or restricted zones.
  • Physical Safety: Avoids hazardous climbing, wildlife interference, or distracting phone use in traffic.
  • Maximum Screen-Free Ratio: Automatically calculates a screen-free target representing 75–85% of the outing's duration.

3. Resilient Architecture & Mock Fallback

  • The FastAPI backend sits between the browser and Ollama as a security barrier, sanitizing untrusted model outputs through Pydantic v2 schemas.
  • If Ollama is offline or warming up, the UI delivers a clear notification rather than fake AI text.
  • For CI/CD and systems without GPUs, a deterministic MOCK_AI=true development mode enables full test suite execution.

4. Client-Side Durability

Active outdoor missions shouldn't be lost if a phone screen locks or a tab reloads while walking. TrailBuddy synchronizes the timer state and active challenge progress with localStorage, while persisting completed sessions permanently to SQLite.

Why Does Open Innovation Matter?

Open innovation and open-weight models are the cornerstone of TrailBuddy's existence:

  1. Zero Surveillance & True Location Privacy: An outdoor companion app frequently touches user schedules, moods, and neighborhood environments. By running open weights locally through Ollama, not a single byte of user context ever touches a third-party cloud server.
  2. No Commercial Paywalls or Token Subscriptions: Closed APIs charge per-token fees and deprecate models without warning. Open-weight models like Qwen3 and Llama give developers permanent ownership and free local inference forever.
  3. Anti-Engagement Philosophy: Proprietary tech platforms build algorithms designed to maximize screen time, ad impressions, and engagement loops. Open innovation gave us the freedom to build software that intentionally measures its success by how quickly users close the application and go outside.
  4. Developer Hackability & Model Agnosticism: Anyone can fork TrailBuddy and swap Qwen3 for Mistral, Gemma 2, or Phi with a single line change in .env (OLLAMA_MODEL=gemma2).

My Agent Session

TrailBuddy was developed in an agentic pair-programming workflow using Google DeepMind Antigravity. The full stack—from database schemas and FastAPI endpoints to the nature-inspired Tailwind design system and automated pytest suites—was built, debugged, and verified end-to-end with automated test runs and live server validation.

Prize Categories

  • Main Track: Hacktoberfest Open-Source AI Challenge Week 1: "Touch Grass"
  • Open-Weight AI: Best use of locally hosted open-weight models (Qwen3 / Llama 3 via Ollama)

Top comments (0)