This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built?
Most people walk through nature completely blind. We step outside with our heads tilted downward at glowing rectangular screens, treating lush canopies, ancient banyans, and migrating birds like passive background wallpaper. Generic "Go outside!" reminders and guilt-inducing streak trackers backfire—they turn the living world into homework.
Nature Connect is a voice-first, mindful outdoor companion created to reverse this disconnect. Its mission is to divert your attention toward one small noticing at a time:
"Listen softly... In the branches above, that gentle hollow tap is the little coppersmith barbet, carving its nest. What a peaceful wonder to notice it today."
Key Highlights:
-Pocket the Screen: Guidance and replies are spoken softly by Leaf, a humble, friendly nature storyteller. You never need to stare at your phone—tap the circle, whisper what you notice, and slide the device back into your pocket.
-Give Things Names: We care about creatures and trees we can name far more than generic scenery. Nature -Connect teaches one name per walk and celebrates personal nicknames ("my mango tree", "the grumpy kingfisher by the bridge").
-Continuity & Long-Term Memory: Love grows from continuity over time. Backboard memory powers genuine callbacks across seasons: "The koel is back! Last year you first welcomed it on March 3rd, so it arrived 9 days earlier this season."
-Humble Storyteller, Never a Dictionary: Leaf speaks with the warmth and gentleness of an elder storyteller sitting under a banyan tree—never clinical Latin names or dry Wikipedia dumps.
-A Real Contributor: Every logged sighting maps onto a community living record. The app celebrates when you are the sole human observer recording a quiet local corner.
-Mindful Keepsakes & 1-Click Social Sharing: Conclude every walk with a keepsake story in your own words, with direct sharing to WhatsApp, Instagram, and Twitter (X) to inspire friends to step outside and touch grass.
Demo
🌐 Live Deployed Web Application on Render: https://nature-connect.onrender.com
🍃 Local & Testing Server: Runs natively on Node.js v20+ with MongoDB and Vite.
The Walk Experience:
The Gate: Walkers sign up with real email validation (featuring live DNS MX domain lookups to block disposable or fake inboxes) and 6-digit verification codes.
The Clearing (/): You enter an unhurried space backdropped by a high-definition nature video. You choose today's mood (e.g. Overwhelmed, Restless, Curious) to receive a tailored 5–15 minute focus plan.
Converse with Leaf: Leaf never interrupts or auto-speaks uninvited. Tap "🌿 Converse with Leaf" to hear a warm greeting. Tapping the embedded glowing leaf button activates speech recognition without awkward holding or premature timeouts.
Quiet Leaf (Instant Mute): Tap "🤫 Quiet Leaf" at any moment to immediately silence spoken playback.
The Keepsake (/recap): View your walk summary, names learned, and contribution to the living map, with instant 1-click sharing buttons for WhatsApp, Instagram Stories, Twitter (X), and direct link copying.
Code
💻 GitHub Repository: https://github.com/icyfire8/Connect-Nature
Architecture Overview
How I Built It
- Open-Source AI Core: Google Gemma (3 / 4B)
The brain of Nature Connect is powered by Google Gemma (running locally via Ollama and deployable on CPU droplets). Gemma performs two critical tasks in our Mastra pipeline:
Semantic Extraction: Parses spoken transcripts into structured entities (speciesHint, nickname, placeLabel, emotion, chatting), understanding informal language like "my big neem tree by the canal" or "a bird going tuk-tuk-tuk".
Humble Storyteller Composition: Rather than dumping robotic taxonomy, Gemma wraps observations in poetic sensory language tailored for either Adult Mode (monsoon seasons, mango blossoms, Sangam migrations) or Kid Mode (nature as a hero's world of brave little characters).
- Mastra Agent Harness
The agent workflow is orchestrated using Mastra (@mastra/core):
stepExtract: Analyzes raw walker transcripts.
stepRecall: Queries MongoDB and Backboard to find if this creature or tree was seen before, calculating days early/late versus last year.
stepEnrich: Fetches live regional context from SerpAPI.
stepCompose: Generates the short spoken audio script and the longer keepsake story.
stepPersist: Saves sightings, nicknames, and assistant memories.
- Voice Output with ElevenLabs
Spoken output is powered by ElevenLabs using warm premade storyteller voices (George - Warm Storyteller, Jessica - Playful & Bright). This ensures Leaf sounds like an affectionate companion resting under the trees rather than a robotic device.
- Continuous Memory with Backboard.io
Every named tree, bird observation, and seasonal milestone is preserved in a dedicated Backboard.io assistant (Nature Connect · [Walker Name]). Even months later, Backboard remembers past encounters to build genuine emotional continuity.
- Local Ecological Facts via SerpAPI
SerpAPI queries the Google Knowledge Graph to enrich observations with real-time seasonal markers—flowering schedules, local monsoon timings, and migratory bird arrival windows across Indian landscapes.
- Cloud Deployment via Render
Configured with an automated render.yaml blueprint. Render builds the static frontend bundle (npm run build) and runs the Node.js Express server on 0.0.0.0:10000.
Why Does Open Innovation Matter?
Privacy for Intimate Outdoor Moments: Walks in nature are deeply personal. Walker observations, GPS coordinates, personal memories, voice transcripts, and children's thoughts should not be locked into an opaque corporate data silo or fed into an advertising model. Open-weight models like Gemma allow local and self-hosted inference where data stays entirely in the user's hands.
Local Ecological Adaptability: Closed, centralized models are heavily biased toward North American geography and temperate fauna. An open-source stack allows us to prompt, fine-tune, and adapt Gemma to local biodiversity—from the coppersmith barbets and koels of Uttar Pradesh to monsoon mango seasons and Sangam river migrations.
Resilience & Swappability: Closed APIs can deprecate features, change pricing, or shut down overnight. With an open-source core (Gemma + Mastra), the foundation of Nature Connect remains permanently functional. Cloud services (ElevenLabs, SerpAPI, Backboard) act as swappable enhancements rather than single points of failure.
My Agent Session
This project was built and iterated pair-programming with Google DeepMind's agentic AI assistant. The agent session helped design the Mastra workflow pipeline, diagnose and eliminate the speech recognition timeout bug, implement DNS MX verification, integrate partner APIs (Gemma, Backboard.io, ElevenLabs, SerpAPI), and prepare the Render deployment configuration.
Prize Categories
This project is a solo submission entered into the following partner tracks:
Render: Deployed on Render as a production Node.js + Vite web service using a dedicated render.yaml blueprint.
ElevenLabs: Powers Leaf's spoken outdoor replies using high-fidelity premade storyteller neural voices (George and Jessica).
Gemma: Utilizes Google's open-weight Gemma model as the core extraction and poetic composition engine.
SerpAPI: Queries the Google Knowledge Graph for real-time regional ecological context, flowering patterns, and migratory bird timelines.
Backboard.io: Powers long-term memory assistants to maintain walker history, personal nicknames, and year-over-year seasonal continuity
Top comments (0)