TrailBurn — Touch Grass, Burn Calories: An Offline-First, Open-Weight AI Outdoor Fitness Advisor

Ashish kumar — Sat, 10 Oct 2026 13:50:19 +0000

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

🌿 TrailBurn — Touch Grass, Burn Calories: An Offline-First, Open-Weight AI Outdoor Fitness Advisor

What I Built

Most modern fitness apps do the exact opposite of what outdoor exercise should be: they keep you staring at a glowing screen, scrutinizing live heart rate graphs, tapping through ads, or following indoor video workouts.

I built TrailBurn around a simple guiding principle: the screen should be the shortest part of the experience.

TrailBurn is an open-source outdoor exercise advisor powered by local open-weight AI that answers a straightforward question:

"I want to burn 400 calories today outside — what can I do, and how long will it take?"

Who Is It For?

Outdoor enthusiasts, hikers, and trail runners who want scientifically accurate calorie goals without needing an internet connection on remote trails.
Privacy-conscious individuals who refuse to upload their body weight, workout habits, and health metrics to closed proprietary cloud servers.
Everyday people looking to touch grass — whether through trail hiking, kayaking, park calisthenics, brisk walking, or even heavy yard work and gardening.

What Does It Do?

Goal-Driven Calorie Calculator: You select your target calorie burn (from 50 kcal up to 2,000+ kcal) and body weight.
Accurate Exercise Duration Solver: Using peer-reviewed MET (Metabolic Equivalent of Task) data from the Ainsworth Compendium of Physical Activities, TrailBurn calculates the exact duration needed across 25+ outdoor activities spanning 5 outdoor environments:
- 🥾 Trails: Easy walks, moderate ascents, steep scrambles, trail runs
- 🌳 Parks: Brisk walking, jogging, outdoor calisthenics, frisbee, yoga, pickup soccer
- 🌊 Water: Leisure swimming, lap workouts, kayaking, paddleboarding, surfing
- 🏙️ Urban Outdoors: Casual biking, high-tempo cycling, roller skating, jump rope, public stair repeats
- 🌻 Garden & Yard: Light gardening, heavy soil work, push mowing, firewood chopping, leaf raking
Local AI Coach: Connects directly to Ollama running local open-weight models (like Llama 3.2, Mistral, Phi-3, or Gemma 2) to stream personalized outdoor workout plans, weather prep tips, hydration guidelines, and warm-ups without a single byte of telemetry leaving your machine.

Demo

GitHub Repository: https://github.com/ashishkr710/TrailBurn
Local Dev Preview: Run locally in seconds via npm run dev at http://localhost:5173/

How It Works in Action

Set your profile: Weight (e.g. 75 kg / 165 lbs), fitness level (Beginner, Intermediate, Advanced), and target calories (e.g. 450 kcal).
Select your outdoor scenery: Choose from all terrains or filter down to Trails, Parks, Water, Urban, or Gardening.
Get instant durations: See immediately that 450 kcal equals:
- ~48 minutes of trail running
- ~1 hour 15 minutes of moderate hiking
- ~1 hour 45 minutes of brisk park walking
- ~1 hour 2 minutes of stand-up paddleboarding
- ~1 hour 30 minutes of heavy garden digging
Chat with your AI Coach: Ask "I'm hiking an unfamiliar trail in the late afternoon. What's a smart pacing and safety plan for burning 500 kcal?" and receive a real-time streaming answer from your local open-weight LLM.
Shut your laptop, head outside, and touch grass!

Code

The complete source code is open source under the MIT License on GitHub:

ashishkr710 / TrailBurn

TrailBurn

View on GitHub

Key Architecture Highlights

src/exercises.js: Contains the full MET database for 25+ activities with beginner, intermediate, and advanced intensity factors, plus pure mathematical formulas solving for duration: $$\text{Duration (hours)} = \frac{\text{Target Calories}}{\text{MET} \times \text{Weight (kg)}}$$
src/ai.js: Client-side gateway connecting directly to local Ollama instances (http://localhost:11434/api/chat). Features automatic model detection, streaming SSE chunk parsing, fallback handling, and structured prompt engineering.
src/style.css: Custom forest-inspired dark aesthetic, fluid glassmorphism, responsive grid layout, and ambient spore/particle canvas animation written in pure Vanilla CSS without heavy UI library dependencies.

How I Built It

The Open-Source AI Stack

Local Open-Weight Models:
- Llama 3.2 (Meta): Tested with both the 1B and 3B parameter open-weight models for rapid, low-latency mobile/laptop inference.
- Mistral 7B & Gemma 2: Supported seamlessly through Ollama's model swap interface for higher-depth outdoor planning.
Local Inference Engine:
- Ollama: Serves open-weight GGUF models directly via a local REST API (/api/tags for model detection and /api/chat with stream: true).
Frontend & Math Foundation:
- Vite + Vanilla JavaScript: Zero bloat, instant startup time, offline cacheability.
- Compendium of Physical Activities: Ainsworth et al. (2011) exercise physiology data embedded directly in the client bundle.

Why Does Open Innovation Matter?

The theme of this week's challenge is Touch Grass with open-source AI at its core. Here is why an open innovation approach is strictly superior to closed proprietary APIs for this project:

1. Zero Cellular Reception on the Trail

When you pack your gear and go outside — deep into a state park, climbing an alpine ridge, or paddling across a mountain lake — cellular reception drops to zero.

A proprietary model behind an API (like OpenAI or Claude Cloud) fails immediately with Connection Refused or Network Unavailable.
Because TrailBurn is built with open-weight models and local inference via Ollama, the AI coach and calculation engine function 100% offline. You can carry your laptop or edge device into a camper van or off-grid cabin with zero signal and still get complete workout planning.

2. Radical Health Privacy (No Corporate Cloud Telemetry)

Your body weight, body composition, fitness level, and workout habits are sensitive personal health metrics.

Closed API services log prompts, retain chat histories, and often train future models on user queries.
With open-source AI, no personal data ever leaves your computer. There are no cookies, no tracking pixels, no telemetry, and no third-party data broker pipelines.

3. Model Autonomy & Swappability

Closed APIs change under your feet: models get deprecated, system prompts get throttled, and pricing changes at will. With open weights:

On a lightweight, battery-efficient ultrabook, you can spin up llama3.2:1b or phi3:mini to conserve wattage on the trail.
On a workstation at home, you can run mistral:7b or llama3.1:8b for detailed multi-day hiking expeditions.
You have complete control over system prompts, temperature, and top-p sampling.

4. Zero Token Inequity & Cost Barriers

Health and physical wellness should not require a monthly software subscription. Closed APIs impose paywalls and token metering. Open-weight models democratize access to personalized, evidence-backed fitness guidance for anyone with a computer at $0 cost.

My Agent Session

This application was developed pair-programming with Google Antigravity / Gemini CLI. The session covered:

Ideating an offline-first fitness concept adhering to the Touch Grass challenge requirements.
Extracting and codifying MET values from exercise physiology research into pure JavaScript functions.
Architecting an asynchronous streaming Ollama gateway that auto-probes local models.
Refining the dark forest glassmorphic UI with smooth slider interactions and responsive layouts.

(Agent session transcript generated with DevRelay / Antigravity IDE).

Prize Categories

Touch Grass Week 1 Main Challenge: Building with open-weight models that get people off the screen and into the real world.
Open-Source Innovation Track: Local inference, privacy-first architecture, and offline trail readiness.

DEV Community: Ashish kumar

[Boost]