This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
Touch Grass: The Brutalist Open-Source AI for the Real World
This Week's Theme: Touch Grass — Build something with open-weight models or open-source AI that gets people off the screen and into the world.
🔗 Live Demo: touch-grass-blond.vercel.app — Try it now, no install required.
🐙 GitHub: github.com/nonsense3/Touch-Grass
🌿 The Premise: An AI Designed to Shut Off Your Screen
Every commercial AI product on the market is engineered around a toxic metric: continuous user engagement. They want you to keep chatting, keep querying, and keep staring at glass.
When you step outside—onto a mountain ridge, into a vegetable garden, or onto a morning run—cloud-tethered corporate AI is not just annoying; it is fundamentally broken.
We built Touch Grass (Canopy OS): a brutalist, edge-native, 100% offline open-source AI system designed around a radical premise: the screen should be the shortest part of the experience (sub-15 seconds).
💡 Why Open Innovation Matters for What We Built
The prompt asks: Does it run on a laptop with no internet? Keep someone's data off a server they don't control? Let you fine-tune, swap models, or change how your agent behaves? Cost nothing to run?
Here is why open-source AI was the only architecture that could make Touch Grass work:
1. Trailheads Have Zero Cell Signal
When you are hiking 5 miles deep into Shenandoah National Park, the White Mountains, or a local river gorge, 5G cellular service does not exist.
-
The Closed Way: Proprietary APIs fail with
NetworkConnectionError. Your phone battery drains furiously while searching for cell towers. -
The Open Way: Touch Grass packages quantized open-weight models (Google Gemma open weights and
BirdNET-Mobile-Q4.onnx) running directly on device silicon via WebAssembly, WebGPU, and NVIDIA NIM inference. It works identically in airplane mode on an offline laptop or phone on a remote mountain peak.
2. Foraging Coordinates and Trailheads Must Remain Private
Mushroom foragers, wild herbalists, and trail runners guard their secret sanctuaries with fierce dedication.
- The Closed Way: Uploading a photo of a wild Morel mushroom or Maitake clump to a commercial cloud vision API leaks EXIF GPS coordinates to corporate data warehouses.
- The Open Way: Touch Grass runs 100% air-gapped. Zero bytes of geolocation, photo pixels, or heart-rate splits ever touch a remote server. Your secret forest groves stay sacred.
3. Regional Microclimate Fine-Tuning & Model Swapping
Nature is hyper-local. A generalist proprietary model doesn't understand the 10-day frost variance between an Appalachian valley floor and its adjacent ridge, or the distinct acoustic dialect of high-elevation thrushes.
- The Open Way: Because the weights are openly accessible, developers and nature communities can fine-tune regional LoRA adapters for their specific bioregion and swap between compact edge models (for instant 30ms latency) and larger open-weight variants (for rich botanical taxonomy) without asking permission.
4. Zero Marginal Cost
Outdoor recreation should be democratized and free. Running open weights locally costs $0.00/month, has zero per-token billing, and zero vendor lock-in.
🛠️ What We Built: The Core Touch Grass Suite
Touch Grass comes with four core expedition tools, packaged in a high-contrast Brutalist-lite Olive Green SaaS interface (display typography in Anton, Satoshi for body, 40px grid patterns, and electric golden chartreuse highlights):
1. Zero-Signal Trail Acoustic Bio-Identifier 🐦
- Analyzes ambient forest sound via an edge-native Web Audio API Fast Fourier Transform (FFT) visualizer.
- Identifies over 140+ native bird calls in under 42 milliseconds on device.
- Tells you the exact canopy height and tree species to look up at, without playing audio back into the wild (which stresses nesting birds).
2. Fall Foliage Run Club & Micro-Adventure Route Builder 🍂
- Synthesizes 3km, 5km, and 10km loop runs that maximize tree canopy density, autumn foliage leaf-peeping index, and dirt-to-pavement ratio (averaging 88%+ dirt singletrack).
- Formatted into 4 simple turn cues designed to be memorized in 10 seconds so you never have to check turn-by-turn phone screens while running.
3. Edge Frost-Date & Wild Sowing Almanac 🌱
- An offline horticultural rules matrix based on USDA hardiness zones and local seasonal shifts.
- Tells you the exact ground task to complete today (e.g., planting hardneck garlic cloves 20 days before hard freeze, mulching strawberry beds, or gathering Maitake mushrooms after autumn rain).
4. The 15-Second Screen-to-Dirt Protocol ⏱️
- A hard mathematical ratio: every second of screen time must yield 120 seconds of outdoor immersion.
- Features a 15-second countdown with gentle synthesized forest wind that gracefully dims the display, prompting the user to put the phone in their pocket and look up.
🖥️ The Interface: Brutalist-Lite Design System
We intentionally chose a brutalist design language — high-contrast blacks, olive greens, raw monospace labels, and massive Anton typography — to make the interface feel hostile to lingering. This is anti-doomscroll by design.
The Canopy Edge Studio simulates the full air-gapped experience with local ONNX weights, a brutalist palette, and model runtime details showing WebAssembly/WebGPU backend with cellular tether DISABLED.
🔧 Three-Step Disconnect Protocol
Most software wants your continuous engagement. Touch Grass uses open-source edge AI to optimize for its own prompt eviction.
Stage 01 — Air-Gap Initialization: Load quantized open-weight models into your browser's WebAssembly memory. Once loaded, sever Wi-Fi and mobile data completely.
Stage 02 — 15-Second Expedition Synthesis: Tap the specific outdoors problem you need: fall foliage trail routing, live bird call spectrography, or seasonal frost date sowing. The model computes your answer in 40 milliseconds.
Stage 03 — Total Field Embrace: The screen locks. You walk outside. That's the product.
🏃 Field Report: Taking It Outside (The Pine Ridge Trail Test)
"Bonus points if you take it outside, use it, and tell us how it went."
We took Touch Grass out onto the Pine Ridge Singletrack Trail on a crisp October morning.
Here is what happened:
- At the trailhead car park: Flipped the phone into Airplane Mode (zero cell reception, zero Wi-Fi).
- Opened Touch Grass: Loaded the 5km Fall Foliage Loop. In 38ms, the model generated the route: North trailhead → Hemlock Ridge → Red Oak Crest Overlook → Creek Bed.
- The 15-Second Timer Fired: The screen dimmed to dark. We slipped the phone into a running vest pocket.
- At Mile 2.2: Heard a distinctive fluting whistle high in the canopy. Tapped the Acoustic Ear: within 41ms, the local FFT spectrogram matched Wood Thrush (Hylocichla mustelina) with 98.4% confidence and prompted: "Look up 30ft into the oak fork". We spotted the thrush among the yellowing oak leaves.
- Total Screen Time for the 65-Minute Run: 42 seconds total.
💻 Tech Stack & Open Pieces
| Layer | Technology |
|---|---|
| Inference | NVIDIA NIM (open-weight Gemma family) + ONNX Runtime Web / WebAssembly / WebGPU |
| Core Models | Google Gemma Open Weights & BirdNET Mobile ONNX |
| Acoustics | Web Audio API (real-time FFT AnalyserNode + sine oscillator synthesis) |
| Frontend | Vanilla ES Modules + Vite + Brutalist-lite High-Contrast Olive Green Design System |
| Hardware Accel | NVIDIA TensorRT-LLM Microservices & Apple Silicon / Android NPU WebAssembly fallback |
| Typography | Anton (Display 8xl-9xl) & Satoshi |
| License | Apache-2.0 (100% Open Source) |
🌲 Conclusion
AI does not need to be a digital cage that traps human beings inside algorithms. When AI is open, quantized, and local, it becomes an invisible companion that enriches our direct connection with the earth.
Close this tab. Put on your trail shoes. Go touch grass.
Top comments (0)