This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
EcoStride AI is a privacy-first, offline-capable mobile companion web app designed to incentivize people to close their laptops, grab their walking shoes, and explore local nature trails.
Unlike traditional fitness apps that focus purely on steps or calories, EcoStride AI acts as a real-world "Druid companion." It utilizes local computer vision to identify plants, birds, and trees on the fly, unlocking lore, achievements, and generating localized audio stories based on the user's immediate surroundings. It’s built for stressed developers, remote workers, and families looking to gamify outdoor exploration without being constantly tracked by ad networks.
Demo
🚀 Live App Deployment: https://vercel.app
📺 Video Walkthrough (YouTube): https://youtube.com
Code
You can explore the full source code and contribute to our open issues here:
github github-vishwajeetCSE/ecostride-ai
How I Built It
EcoStride AI is built completely on top of open-source, local-first architectures to make it resilient even deep in the woods with zero cellular service:
Local Vision Inference: We use Transformers.js running a quantized MobileNetV4 directly in the browser web worker to handle real-time plant and wildlife identification locally on the device.
Agent Orchestration: The app uses LlamaIndex locally to manage a Retrieval-Augmented Generation (RAG) system containing local regional flora/fauna encyclopedias.
Local LLM Execution: For offline audio-adventure generation, we leverage Ollama running locally on the desktop companion app, or fallback to an open-weights Llama-3-8B-Instruct instance hosted transparently via Hugging Face Inference Endpoints when online.
Why Does Open Innovation Matter?
Open innovation completely transformed what was possible for this application. A project like EcoStride AI inherently demands extreme user privacy—nobody wants their real-time hiking locations and camera feeds sent to a proprietary black-box API.
By building on open-weights models and local inference tools, we ensure that user data never leaves their device. Furthermore, using open-weights models meant we could fine-tune and quantize the wildlife classifier specifically for edge devices without paying prohibitive per-API-call fees, allowing us to keep this software entirely free, accessible, and community-driven.
My Agent Session
I used DevRelay to record the multi-turn agent system parsing the visual input and dynamically updating the user's quest log. You can see the full internal logic and tracing here:
agent_session devrelay-session-id-placeholder
Prize Categories
Most Creative Use of Local-First AI
Best Open-Weights Integration (Hugging Face / Llama 3)
Top comments (0)