This is a submission for the MLH x DEV Writing Challenge
What I Built
For the Hacktoberfest Week 1 "Touch Grass" challenge, I built TerraAgent, an open-source agentic workflow designed to generate hyper-local, weather-aware outdoor itineraries. Instead of doom-scrolling, TerraAgent acts as an outdoor copilot, discovering nature trails, parks, and botanical spots specifically optimized for your immediate local area.
The core of the project relies on open-weight AI. I utilized an open Llama 3.3 model for inference, orchestrated with LangChain and Pydantic to ensure the agent outputs strictly structured, actionable outdoor plans rather than conversational fluff.
Open innovation is critical for a project like this. Outdoor data—like exact trail coordinates, local foraging spots, and off-grid routing—should remain in the hands of the community, not walled off on a proprietary server you don't control. By building on open-weight models, the architecture is designed so that inference can eventually be run entirely locally on a mobile device, keeping hikers disconnected from the cloud but fully aware of their surroundings.
Demo
Here is a design prototype of TerraAgent in action on the trail. By leveraging open-weight models, it functions entirely offline, recommending hyper-local routes like the Ananthagiri Forest Walk in Hyderabad and identifying point-of-interest markers without needing a cellular connection.
The interface is built to make interactions as brief and high-utility as possible: hikers can generate an offline route map, see localized markers like "Local Vegetation," and then quickly tuck the phone away to enjoy the outdoors.
Partner Technologies
I deployed the backend FastAPI service on Render. Having a robust platform to host the Python logic and agent endpoints made the deployment seamless, allowing me to focus entirely on the orchestration logic rather than fighting with infrastructure.
I also leveraged GitHub Copilot heavily while scaffolding the Next.js and Tailwind CSS frontend, using it as a true pair programmer to rapidly build out the interactive UI components so I could spend more time refining the AI's tool-calling accuracy.
Hackathon Experience
Building this felt like a direct continuation of the momentum from MLH Global Hack Week: Agents. Diving deep into autonomous loops, RAG, and tool calling during GHW perfectly laid the groundwork for TerraAgent. Competing alongside other builders in the agentic space has completely shifted how I view production architecture—moving from static web apps to dynamic, environment-aware copilots that actually encourage you to close your laptop and step outside.
Top comments (0)