DEV Community

Cover image for TouchGrass: The Open-AI Agent That Succeeds When You Stop Using It
Rajan Mishra
Rajan Mishra

Posted on

TouchGrass: The Open-AI Agent That Succeeds When You Stop Using It

#devchallenge #hf26challenge #opensource #ai

Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass Submission 🌿

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

What I Built

Most modern consumer software is built around a single, predatory funnel:

$$\text{Attention} \longrightarrow \text{Screen Time} \longrightarrow \text{Algorithmic Feed} \longrightarrow \text{Ad Impressions}$$

Even modern AI tools often optimize to keep you typing in a chat window forever.

TouchGrass reverses this equation.

TouchGrass is an open-AI-powered real-world challenge platform built on the principle:

The application succeeds when the user stops using it.

Instead of outputting text for you to consume on a screen, TouchGrass's primary output is real-world physical action.

The system uses an open-weight Gemma 2 agent orchestrator to understand your personal context, difficulty level, and history, then generates personalized missions across 9 real-world categories:

  1. Outdoor: Horizon walks, sun & shadow tracking, reverse-loop neighborhood walks.
  2. Nature: Flora texture expeditions, wild seed nurturing, avian song identification.
  3. Kindness: Anonymous door holding, carrying neighbor groceries, sincere compliments.
  4. Happiness: Daylight resets, ripple smiling.
  5. Human Connection: Screen-free coffees, listening to elderly neighbors.
  6. Community: Sidewalk plastic cleanups, local library contributions.
  7. Health: Power walks, barefoot lawn grounding, outdoor mobility stretches.
  8. Creativity: Urban texture discovery.
  9. Exploration: Turning down unexplored alleyways.

The Behavioral Rhythm: Attract âž” Release âž” Reward

  1. State A — Attract: A crisp, inspiring mission briefing with XP bounties and an ElevenLabs voice narration.
  2. State B — Release (Zen Mode): The moment you accept, the UI visually disappears into a minimal ambient breathing circle: > YOU'RE OUTSIDE. GO DO IT. You don't need this screen. A timer logs your outdoor minutes while your phone stays in your pocket.
  3. State C — Reward: When you return and speak your reflection ("I helped my neighbor carry her groceries"), the Gemma Evaluator scores your sincerity, awards bonus XP, triggers level-up celebrations, and unlocks new features.

Demo

Code & Deployment

The complete codebase is 100% open-source under the MIT license on GitHub:
👉 github.com/aiwithrajan/touchgrass

  • Hosting & Runtime (Render): Deployed as a unified web service on Render, serving both the FastAPI agent backend and high-performance React SPA.
  • Backend: FastAPI + Gemma 2 Agent Orchestrator + Safety & Diversity Policy Engine + Local-First Memory Store.
  • Microclimate AI: TabPFN Tabular Foundation Model (Prior Labs) for optimal outdoor window and vitality scoring.
  • Frontend: React 19 + TypeScript + Tailwind CSS v4 (Calm, distraction-free flat editorial UI system).
  • Voice: ElevenLabs API with on-device Web Speech API fallback.
  • Observability: Sentry Agent Tracing instrumentation.

How I Built It

1. Gemma 2 Open-Weight Agent Orchestrator

At the core of TouchGrass is a closed-loop real-world agent. When you request a mission, Gemma considers your level, your recent category history, and time constraints.

To prevent the common LLM trap of repeating the same task ("Go for a walk, go for a walk, go for a walk"), the agent pipeline feeds into our Policy Engine:

  • Safety Policy: Scans candidate missions for dangerous activities, trespassing, risky maneuvers, or confrontation.
  • Diversity Policy: Enforces category rotation so nature is followed by kindness, then community, then creativity.
  • Progression Policy: Gates advanced human/community challenges until the user establishes personal grounding in Levels 1–3.

2. The Voice Layer

Using ElevenLabs (with Web Speech API fallback), Gemma's mission is spoken aloud in a calm, encouraging cadence:

"Hey. Your challenge today is simple. Go outside, look up, and give a genuine smile to three people. I'll see you when you're back."
This allows you to absorb the mission through headphones or speaker and immediately put your phone in your pocket.

3. Voice Reflection & Gemma Evaluator

When you return, you tap the mic and speak your reflection. The Gemma Evaluator analyzes the sincerity and depth of your outdoor effort, awarding bonus XP (+10 to +50 XP).

4. Progression & Milestones

  • Level 5 Unlocks Social Mode: A privacy-first community board showing what other verified humans did outside (only approximate city area, zero GPS tracking).
  • Level 10 Unlocks Journey Sharing: A canvas/SVG card ready to export and share with your real-world achievements.

5. TabPFN Microclimate Vitality Forecaster

Going outside is most effective when synchronized with optimal daylight and comfort. Using TabPFN (Prior Labs Tabular Foundation Model), TouchGrass analyzes hourly solar elevation, temperature gradients, ambient humidity, and UV levels to predict:

  • Optimal Outdoor Window: Calculates the exact hour (e.g. Today 10:30 AM) that maximizes natural circadian daylight while avoiding harsh midday UV glare.
  • Vitality Score (0–100): A continuous rating balancing atmospheric freshness, light spectrum, and seasonal conditions.
  • Offline Resilient: Operates with a 100% on-device tabular prior engine, guaranteeing backcountry functionality even when cell towers are miles away.

Why Does Open Innovation Matter?

This was the central question of Week 1, and it is the exact reason TouchGrass was conceived:

1. Zero Signal in the Backcountry (Offline Trail Capability)

When you hike a ridge trail, walk through a national park, or step into a forest, cellular internet disappears.
Closed proprietary APIs (like commercial OpenAI or Anthropic endpoints) fail instantly without internet connectivity. Because Gemma and TabPFN run locally on-device, TouchGrass continues to forecast microclimates, generate challenges, and evaluate reflections even with your phone in Airplane Mode.

2. Complete Location & Behavioral Privacy

Tracking daily personal reflections, outdoor routines, and physical habits shouldn't require sending your life's diary to third-party ad networks. With open-weight AI and local-first storage, your thoughts and habits stay completely on your device.

3. $0 Cost to Run

There are no recurring token subscription fees or rate limits. Anyone with an entry-level laptop or smartphone can run the engine forever.

Prize Categories

  • Overall Grand Prize ($250): Open-weight AI with Gemma 2 at its core, reversing screen addiction into physical real-world presence.
  • Best Use of Render ($200): Hosted on Render as a unified full-stack AI agent runtime, serving both the FastAPI backend and React SPA with automated zero-downtime builds.
  • Best Use of TabPFN ($200): Prior Labs Tabular Foundation Model forecasting the optimal 15-minute outdoor window and vitality score from microclimate tabular data.
  • Best Use of ElevenLabs ($200): Calm, spoken voice mission briefings and audio playback enabling screen-free immersion.
  • Best Use of Sentry ($200): Comprehensive agent tracing instrumenting Gemma generation, TabPFN tabular prediction spans, and policy checks.

Taking It Outside: Field Test

I took TouchGrass outside on an afternoon loop. The agent gave me:

"Avian Song Tracker: Sit outside in silence until you can pinpoint the physical location of at least two singing birds. Watch their flight path without taking a photo."

I accepted, the screen went completely black into Zen mode, and I put the phone in my jacket pocket. I spent 22 minutes under an oak tree, heard a blue jay and two sparrows, and didn't touch my phone once.

When I got home, I opened the app, saved my reflection note into my personal Journey diary, and watched my 7-day activity tracker mark the day complete.

Real life > Screen time.

Top comments (0)