SIDEQUEST: An Open-Source AI Outdoor Exploration Game (Touch Grass with In-Browser Sentence Transformers)

Om Honrao — Sat, 10 Oct 2026 09:52:09 +0000

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

What I Built

Most AI applications trap our attention behind glowing glass. We built SIDEQUEST with the opposite philosophy:

"The screen should help someone enter the world, not compete with it."

SIDEQUEST is an open-source outdoor exploration web game that transforms ordinary surroundings—college campuses, neighborhood alleys, apartment balconies, and small gardens—into personalized investigations.

Instead of tracking steps, badges, or endless map feeds, SIDEQUEST creates an evolving observation loop:

Set Intent: Tell the local AI how much time you have (5, 10, 15, or 30 minutes), your mood (Curious, Calm, Adventurous, Scientific, Creative, Observant), and your environment.
Local AI Mission Matching: An open-weight sentence-embedding model running 100% locally in your browser scores 36 curated investigations in real time (~12ms) using cosine similarity.
Pocket Mode: Read or listen to the mission briefing and safety notes once, then lock your phone and put it in your pocket.
Physical Exploration: Search for subtle phenomena (e.g., plants thriving in concrete fissures, thermal gradients between building shadows, 360° acoustic horizons, ant highway seam navigation).
Field Journal: Return and log your raw observations, reflections, and optional local photos.
Local Discovery Engine: The local sentence transformer embeds your field note, identifies the underlying ecological phenomenon (e.g., botanical resilience, microclimatic thermal mass, acoustic masking), and generates an editorial Discovery Card with follow-up curiosity questions and a personalized next mission.

It is designed for students, developers, urban naturalists, and anyone experiencing screen fatigue who wants to cultivate deep attention without traveling to remote national parks or needing expensive gear.

Demo

Repository: https://github.com/1nv1sibl3/Hacktober-Week1-SIDEQUEST
Local Startup: Clone the repository and run npm run dev to launch locally in under 1 second.
Workflow in 60 Seconds:
1. Click "Begin New Expedition" → Select 10 min, Curious, Campus, and type "plants growing in cracks".
2. The local model ranks "The Survivor" as #1 in ~12ms.
3. Listen to the audio briefing, tap "Enter Pocket Mode", lock your screen, and explore.
4. Tap "I'm Back" upon return, record your observation, and receive your synthesized Discovery Card and recommended follow-up quest.

Code

1nv1sibl3 / Hacktober-Week1-SIDEQUEST

🌿 SIDEQUEST — An Open-Source AI Outdoor Exploration Game

Built for the MLH / Hacktoberfest 2026 Open-Source AI Challenge, Week 1: "Touch Grass".

SIDEQUEST transforms ordinary outdoor environments into personalized exploration games. The user receives an investigation tailored to their time and surroundings, leaves their phone in their pocket, explores the physical world, and returns to log what they discovered. An open-weight sentence-embedding AI model running 100% locally in the browser connects their field notes to ecological insights and selects increasingly relevant follow-up missions.

Core philosophy: The screen should help someone enter the world, not compete with it.

🌟 Key Differentiators

Local Semantic Intelligence: Computes real 384-dimensional sentence embeddings in-browser via Transformers.js (all-MiniLM-L6-v2, Apache-2.0). Zero cloud API latency or fees.
Zero-Cloud Core Experience: Works completely offline after initial model weight download. No account required, no remote databases, zero tracking telemetry.
Observation-Driven Mission Evolution: Your field notes…

View on GitHub

Key Architectural Highlights:

Zero Cloud Egress: 100% of user field notes, photos, and milestones persist in client-side IndexedDB.
Inspectable AI Lab: Includes a dedicated model diagnostics dashboard displaying live vector previews (384-dimensional normalized floats), inference latency, and cosine similarity rankings.
Hardware-Resilient Pocket Timer: Mobile OS battery managers suspend background JavaScript timers when devices lock. SIDEQUEST calculates elapsed duration against hardware clock delta timestamps (Date.now() - startedAt) upon wake, completely eliminating timer drift.

How I Built It

The Open-Source AI Stack

Open-Weight Model: sentence-transformers/all-MiniLM-L6-v2 by Microsoft & Hugging Face, distributed under the Apache-2.0 license.
Inference Runtime: Powered by Transformers.js (@xenova/transformers) running quantized ONNX models directly in the browser over WebAssembly (WASM SIMD) and ONNX Runtime Web.
Zero-Latency Vector Search: All 36 curated missions ship with pre-computed 384-dimensional normalized vectors, enabling instant similarity rankings on first launch.
Transparent Deterministic Fallback Scorer: If WebAssembly or network downloads fail prior to initial cache creation, a built-in TF-IDF vector index and category scorer activates with zero disruption. We never fake AI or simulate random numbers.
Audio & Accessibility: Web Audio API zero-asset harmonic bell synthesizer (528 Hz chime) and Web Speech API hands-free screenless read-aloud.
Frontend Core: React 19, TypeScript, Tailwind CSS v4, Lucide Icons, and Canvas Confetti.

Why Does Open Innovation Matter?

For an outdoor exploration product, open-source, open-weight AI is fundamentally superior to closed proprietary cloud APIs:

True Disconnected Offline Resilience: Closed APIs (OpenAI, Gemini, Claude) fail the moment you walk under thick tree canopies, enter urban canyons, or hike off-grid where cellular reception drops. An open-weight model cached in the browser's Cache API runs completely offline in airplane mode.
Absolute Personal Privacy: Nature journals and personal reflections are intimate. With open-weight local inference, zero user data, location tags, or photos ever leave the browser.
Permanent Zero Cost & No Rate Quotas: You never have to worry about monthly API billing, credit card expirations, or subscription lockouts while on an expedition.
Mathematical Verifiability: In our AI Lab, users can inspect the exact cosine similarity math and raw float vectors rather than relying on black-box proprietary hallucinations.

My Agent Session

This application was conceptualized, architected, implemented, and verified using Google Antigravity paired with open-source development tooling. The entire primary user journey—from preference weighting and local ONNX embedding execution to IndexedDB session persistence—was validated with automated Vitest integration test suites.

Prize Categories

Hacktoberfest 2026 Open-Source AI Challenge — Week 1: "Touch Grass"
Most Creative Use of Open-Source / Open-Weight AI
Best Client-Side / Offline-First Experience

DEV Community: Om Honrao