TerraAgent: Touching Grass with an Open-Weight Agentic Copilot

Nalla Sumang — Fri, 09 Oct 2026 07:32:35 +0000

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.

A prototype AI agent can survive on a static CSV and a single Python script. A production agent cannot.

Nalla Sumang — Fri, 09 Oct 2026 07:23:22 +0000

I recently completed the AI-Assisted Data Science with BigQuery lab via Google Cloud. While analyzing the final multimodal vector search stage, I focused heavily on how the pipeline enforces state dependencies.
If the remote multimodal_embedding_model isn't strictly instantiated within the dataset first, the system throws a hard blocker. In an isolated sandbox, that’s just a minor execution step. But in a production multi-agent system? That is a catastrophic DAG dependency failure. This constraint perfectly exposes the exact bottleneck the AI industry is hitting today.
We spend all our time building the orchestration loops, and almost zero time engineering the rigid data pipelines required to feed them. Here is the unfiltered engineering reality of scaling agents:
→ Data Gravity over APIs: Pulling massive datasets out of a warehouse just to generate vector embeddings kills real-time execution. Running the embedding model natively inside the warehouse cuts extraction latency to zero.
→ Strict State Management: You can build the most elegant LangGraph orchestration in the world, but if your agent fires before your remote model and vector indices are fully instantiated, your workflow is dead on arrival.
→ Retrieval Speed = Agent Speed: If your agent has to wait on three external scripts to normalize a multimodal table before it can execute an action, your agent isn't autonomous. It's just a slow API call. Good AI agents don't just reason well. They are plumbed well.

DEV Community: Nalla Sumang

TerraAgent: Touching Grass with an Open-Weight Agentic Copilot

What I Built

Demo

Partner Technologies

Hackathon Experience

A prototype AI agent can survive on a static CSV and a single Python script. A production agent cannot.