DEV Community: Pawan Shivaji ghule

GrassWhisperer 🌿 — A Local AI Outdoor Companion That Works With No Signal

Pawan Shivaji ghule — Sun, 11 Oct 2026 06:30:36 +0000

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

What I Built

I built GrassWhisperer 🌿 — a local-first AI outdoor companion that gets you off the screen and into the world in under 30 seconds.

It's a beautifully designed web app with four AI-powered tools, all running 100% on your laptop with no internet connection, no API keys, and no accounts:

Tool	What It Does
🐦 Bird Whisperer	Describe what you see or hear on the trail → instant species ID, call sounds, field tips
🌱 Garden Oracle	Your city + USDA zone → personalized weekly planting calendar with local frost dates
🥾 Trail Finder	Location + available time → seasonal trail picks (fall foliage, wildflowers, sunrise) + gear list
🤔 Ask Nature	Free-form outdoor Q&A — foraging, stargazing, animal tracking, gear, safety

Who is it for? Hikers, birders, gardeners, runners, and anyone who wants to feel more connected to the natural world. The app is built around a single principle: the screen time should be the shortest part of the experience. Open it, get your answer, close it, go outside.

Demo

🚀 Run it locally in 3 commands:

ollama pull llama3.2:1b          # one-time model download (~700MB)
git clone https://github.com/PawanTheGod/grasswhisperer
cd grasswhisperer && npm install && npm start
# → open http://localhost:3333

App in action:

The Bird Whisperer tab — describe a bird from the field and get a full identification with call sounds and ID tips:

Type: "small bird, bright yellow belly, black cap, white cheeks, pine tree, making 'chickadee-dee-dee' call"
→ Identifies: Black-capped Chickadee with habitat, call description, and field tips

The Garden Oracle — enter your city and get a planting calendar for this exact week, with frost dates pulled from the model's training data:

Type: "Portland, Oregon, Zone 8b"
→ Get: garlic to plant now, frost date info, soil prep task, and what to wait on until spring

Code

PawanTheGod / grasswhisperer

🌿 Local AI outdoor companion — bird ID, garden planner, trail finder. Powered by Ollama + llama3.2. No cloud, no API keys. Hacktoberfest Touch Grass 2024.

🌿 GrassWhisperer

Hacktoberfest 2024 · Challenge 1: Touch Grass
Local AI outdoor companion — 100% open-source, runs on your laptop, no internet needed.

What It Does

GrassWhisperer is a local-first AI app that helps you get off the screen and into the world:

Feature	Description
🐦 Bird Whisperer	Describe what you see or hear — get an instant ID with field tips
🌱 Garden Oracle	Enter your city/zone, get a personalized weekly planting calendar
🥾 Trail Finder	Find the best hike near you for the season (fall foliage, wildflowers, etc.)
🤔 Ask Nature	Free-form Q&A — foraging, stargazing, wildlife tracking, gear

Why Open Source Matters Here

This runs entirely on your laptop using Ollama + llama3.2:1b — a tiny open-weight model that fits in under 1GB of RAM.

✅ Works offline on the trail — no cell signal needed
✅ Your data never leaves your device — no API keys…

View on GitHub

📂 Repository: github.com/PawanTheGod/grasswhisperer

Key files:

server.js — Express server + all 4 Ollama AI endpoints
public/index.html — Full UI with tab navigation
public/style.css — Dark forest theme with micro-animations
public/app.js — Vanilla JS, zero frameworks
setup.js — Auto-setup: checks Ollama, pulls model

Tech Stack:

Runtime  → Ollama (MIT license) — local LLM server
Model    → llama3.2:1b (Meta open weights) — ~700MB, <1GB RAM
Server   → Node.js + Express
Frontend → Vanilla HTML / CSS / JS  (no build step, no bundler)

How I Built It

Architecture

Browser → Express (localhost:3333) → Ollama API (localhost:11434) → llama3.2:1b
                                              ↑
                                 Runs entirely on your machine.
                                 No packets leave your network.

Open-Source AI Core: Ollama + llama3.2:1b

Ollama is the open-source runtime that makes all of this possible. It manages model downloads, runs a local REST API server (identical interface to OpenAI), and handles inference — all through a dead-simple http://localhost:11434 endpoint.

llama3.2:1b is Meta's smallest open-weight Llama 3 model. At ~700MB and under 1GB RAM it runs comfortably on a 4-year-old laptop. Fast enough for real-time Q&A (5–15 seconds per response on CPU).

Prompt Engineering for Structured Output

The key challenge with compact local models is reliable JSON output. I solved this with tight, schema-pinned system prompts:

// server.js — Bird Whisperer system prompt
const system = `You are a friendly field ornithologist.
Respond in this EXACT JSON format (no markdown):
{
  "name": "Common Name",
  "scientific": "Scientific name",
  "confidence": "high/medium/low",
  "calls": "What its call sounds like",
  "habitat": "Where to find it",
  "funFact": "One surprising fact",
  "look_for": "Key identification tips outdoors"
}`;

Then the server extracts JSON safely using raw.match(/\{[\s\S]*\}/) — handling cases where the model wraps its response in extra prose.

Swapping Models is One Line

// In server.js — change this to any Ollama model:
const MODEL = 'llama3.2:1b';   // default — tiny, fast, any laptop
const MODEL = 'llama3.2:3b';   // smarter, better JSON reliability
const MODEL = 'mistral';        // excellent at instruction following
const MODEL = 'phi3:mini';      // Microsoft's ultralight model

Why Does Open Innovation Matter?

A bird ID app that needs the cloud to work doesn't work when you need it most — standing in the woods with no signal watching an unfamiliar bird disappear into the trees.

Here's exactly where open-source AI beat every closed alternative for this use case:

1. 📶 It works offline on the trail.
Once llama3.2:1b is downloaded (one time, 700MB), GrassWhisperer works with zero connectivity. No OpenAI API key. No Gemini quota. No network at all. That's a genuine difference in the real world.

2. 🔒 Your data never leaves your machine.
Where you hike, what birds you're seeing, your garden's location, questions about foraging near your home — none of that touches a server. With closed APIs, every question is logged and potentially used for training. With Ollama + open weights, it stays in RAM and disappears when you close the tab.

3. 💸 It costs nothing to run.
Closed APIs charge per token. Heavy outdoor Q&A use could easily add up to $10–20/month. GrassWhisperer has zero marginal cost after setup — no metering, no billing, no surprises.

4. 🔧 You can swap, extend, or fine-tune the model.
Want a model fine-tuned on bird taxonomy? Swap it in — one line change. Want to add plant identification using a vision model like llava? It's one new endpoint. Closed APIs lock you into their update schedule, their capabilities, and their pricing. Open weights give you full sovereignty.

5. 💻 It runs on cheap hardware.
llama3.2:1b uses under 1GB RAM and responds in under 15 seconds on a CPU. A $35 Raspberry Pi 5 could run this at a community garden, a nature center, or a trail head kiosk. That kind of democratization only happens with open weights.

The screen should be the shortest part of a Touch Grass experience. Open-source AI makes that possible by putting the intelligence on the device in your pocket — not in a data center you're permanently paying to rent access to.

My Agent Session

This project was built collaboratively with an AI coding agent (Antigravity IDE). The agent:

Scaffolded the full Express + Ollama integration from scratch
Designed the prompt engineering strategy for extracting structured JSON from compact models
Built the dark forest-themed CSS with ambient animations and glassmorphism cards
Wrote the auto-setup script to detect Ollama status and pull models
Built the full frontend — tabs, result rendering, error states, example chips

The entire project went from zero to running in a single session — made possible by the simplicity of the open-source stack. No API key management. No cloud deployment. No billing dashboard. Just npm start and it works.

Prize Categories

Best Use of Ollama / Local Inference: GrassWhisperer is built entirely around Ollama as its open-source inference runtime. Without Ollama, the app doesn't exist — it's not an afterthought, it's the foundation. Every AI feature runs through the local Ollama API with zero cloud dependencies.

🐾 Sanctuary: Building a Living AI Companion with Gemma

Pawan Shivaji ghule — Mon, 05 Oct 2026 06:24:27 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

I built Sanctuary — a living AI companion featuring a virtual roommate pet named Mochi, powered by Google's open-weight Gemma 2B model and optional ElevenLabs voice synthesis.

Instead of being another chatbot that waits for a prompt inside a text box, Mochi is an autonomous creature designed to share your digital workspace.

🐾 Continuous Autonomous Life Simulation: Mochi has an ongoing internal state model including energy, happiness, curiosity, restfulness, boredom, and social need. It wanders, stretches, loafs, naps, and reacts to the user without continuously calling an AI model.
👁️ Passive Workspace Awareness: Mochi tracks cursor position, movement, proximity, and user activity. When the user is actively typing, it suppresses unnecessary interruptions and gives them space.
🧠 Gemma-Powered Reflection: When the user wants to talk or needs guidance, Mochi can activate Google's open-weight Gemma 2B model for CBT-inspired reflection. It can help identify patterns such as catastrophizing or all-or-nothing thinking and turn them into more balanced perspectives and practical grounding steps.
🗣️ Warm Voice Interaction: Mochi can speak using ElevenLabs Flash v2.5, with the browser's Web Speech API available as an alternative for users who want a local/offline voice option.
🪟 Floating Companion Mode: Mochi can move around the Sanctuary viewport and can be popped out into a browser-managed floating window using the Document Picture-in-Picture API.
🫁 Grounding Tools: Sanctuary includes guided breathing and ambient calming experiences, including procedural purring and rain sounds.

I built Sanctuary for my friend Alex, a software engineer who struggles with late-night burnout, insomnia, and spiraling self-criticism during difficult work periods.

The problem I wanted to solve was simple:

When someone is overwhelmed, why should they have to open a blank chatbot and figure out what to type?

Mochi is designed to provide quiet digital companionship first. It can simply sit nearby, wander around, take a nap, or react to what you're doing. When you actually want to talk, it becomes an intelligent conversational companion.

Sanctuary is intended as an emotional-support and mindful-grounding companion, not a clinical treatment or replacement for professional therapy.

Demo

🚀 Live Web App: https://sanctuary-0ib8.onrender.com/

In supported Chromium browsers such as Chrome and Edge, use the Pop-out PiP button to detach Mochi into a browser-managed floating companion window while you continue working.

Code

PawanTheGod / sanctuary

🌱 Sanctuary — Living AI Roommate Pet & Supportive Companion

Sanctuary is a supportive companion and living roommate pet ("Mochi") built for a friend navigating burnout, late-night insomnia, and acute anxiety spirals.

(Note: Sanctuary is an emotional support and mindful grounding companion, not a clinical treatment or replacement for professional therapy.)

✨ What Is Sanctuary?

Sanctuary is built around Mochi — an autonomous creature that inhabits your digital workspace.

Unlike traditional chatbot assistants that wait passively for text prompts or send intrusive pop-up notifications, Mochi runs a continuous 20Hz local life simulation:

🐾 Living Presence: Wanders across your screen, stretches, loafs, sleeps with floating Zzz particles, and gets playful micro-zoomies.
👁️ Cursor Awareness: Dynamic eye pupil tracking and proximity dwell detection relative to your cursor.
🧠 Open-Source Gemma Brain: Leverages Google's open-weight Gemma LLM (running on-device via WebGPU / WebLLM or local Ollama) for CBT-inspired cognitive distortion reflection and…

View on GitHub

📂 GitHub Repository: https://github.com/PawanTheGod/sanctuary

Tech Stack:

Vanilla JavaScript (ES Modules)
HTML5
CSS3
Web Audio API
WebGPU
WebLLM (@mlc-ai/web-llm)
Google Gemma 2B
Ollama
ElevenLabs API
Document Picture-in-Picture API

How I Built It

Sanctuary is built around a decoupled architecture where Mochi's body and cognition are separate systems.

1. MochiLifeEngine — The Body

The core mochi-engine.js runs a continuous local life simulation.

Instead of asking an LLM what Mochi should do every few milliseconds, the engine maintains internal biological-style variables such as:

energy
happiness
curiosity
boredom
sleepiness
socialNeed

A priority arbitrator uses these values and the current context to select behaviors such as wandering, approaching the cursor, playing, loafing, stretching, affection, or sleeping.

This means Mochi can feel alive without continuously consuming LLM tokens.

2. ContextProvider — The Senses

context-provider.js acts as Mochi's sensory layer.

It observes local interaction signals such as:

Cursor position and movement
Cursor proximity to Mochi
User activity and inactivity
Typing activity
Browser focus
Session duration
User interaction with Mochi

These events allow Mochi to respond to its environment without requiring an LLM for every decision.

One important design rule is that active typing suppresses unnecessary proactive interruptions.

3. Gemma — The Cognitive Layer

When an interaction actually requires reasoning, Sanctuary escalates from the local life engine to Gemma.

The primary local path uses:

Gemma 2B → WebLLM → WebGPU

with the model:

gemma-2-2b-it-q4f16_1-MLC

Sanctuary also supports a local Ollama path using:

gemma2:2b

Gemma is therefore not responsible for Mochi's continuous physical behavior.

Instead:

Mochi's local life engine → meaningful interaction → Gemma reasoning → Mochi response

This separation was one of the most important architectural decisions in the project.

4. CBT-Inspired Reflection

When the user shares a worry, Mochi can use Gemma to structure the conversation around CBT-inspired reflection.

For example, it can help identify patterns such as:

Catastrophizing
All-or-nothing thinking
Mind reading
Negative assumptions

It then helps turn the thought into a more balanced perspective and a small practical grounding action.

The goal is not to replace therapy, but to make supportive reflection feel more natural and accessible.

5. Voice Layer

For users who want spoken interaction, Sanctuary can route Mochi's responses through ElevenLabs Flash v2.5.

The browser's Web Speech API provides an alternative voice path, allowing the core companion experience to remain usable without requiring ElevenLabs.

6. Floating Companion Surface

Mochi was originally designed as a website companion, but I wanted the pet itself to feel more persistent.

Sanctuary therefore includes a dedicated floating companion layer and a Document Picture-in-Picture surface.

This allows Mochi to leave the main Sanctuary interface and remain visible in a browser-managed floating window while the user works elsewhere.

It is deliberately browser-based rather than pretending to provide a native OS-wide overlay.

Why Does Open Innovation Matter?

Open innovation made an important part of Sanctuary possible: local AI cognition.

Instead of designing Sanctuary around a mandatory cloud LLM API, I could build the cognitive layer around an open-weight model that can run locally.

With Gemma 2B running through WebLLM/WebGPU or Ollama, users can run the reasoning component on their own machine.

That changes the architecture significantly.

The continuous life simulation doesn't need a cloud service at all. Mochi's behavior engine, state, context processing, and interaction logic can operate locally, while Gemma can also run locally when the user chooses a local inference path.

This makes the companion:

Less dependent on network availability
More responsive to local interactions
More controllable by the developer
Better suited to privacy-sensitive conversations when local inference is used
Free from per-token cloud LLM costs in local mode

Open-weight AI also allowed me to experiment with an architecture where the LLM is not the entire product.

Gemma is the cognitive layer.

The rest of Mochi — its body, senses, behavior, memory/state, and personality — is software I designed around it.

That was the part that made Sanctuary possible as more than another AI chat interface.

My Agent Session

I built Sanctuary through an intensive pair-programming workflow with an autonomous AI coding agent.

The agent helped scaffold and iterate on:

Mochi's life simulation
ContextProvider
Behavior prioritization
Gemma/WebLLM integration
Ollama fallback
Voice integration
Floating companion architecture
Document Picture-in-Picture support
Security and secret-leak checks
Documentation and architecture files

The development process was iterative: build → run → test → inspect → harden → repeat.

The important architectural constraint throughout the process was to keep the 20Hz life engine independent from Gemma, so the companion would remain responsive and autonomous rather than turning every movement into an LLM request.

Prize Categories

🏆 Best Use of Gemma

Gemma is the cognitive core of Sanctuary.

The project uses Google's open-weight Gemma 2B locally through WebLLM/WebGPU, with Ollama providing another local inference path.

Rather than using Gemma simply as a chatbot, I integrated it into a larger autonomous companion architecture where it is selectively activated when meaningful reasoning is required.

🔊 Best Use of ElevenLabs

ElevenLabs provides Mochi's expressive voice layer through ElevenLabs Flash v2.5.

The voice is designed to make Mochi feel more like a companion than a conventional text-based assistant, while Web Speech provides an alternative voice path.

🟣 Best Use of Render

Sanctuary is deployed on Render, providing a publicly accessible live application that anyone can try without setting up the project locally.

🚀 Live Demo: https://sanctuary-0ib8.onrender.com/