GeoFire-Agent

Orbital Edge Compute Agent — Autonomous Satellite Triage & Fire-Risk Platform

Overview

GeoFire-Agent has evolved from a deterministic PostGIS web-app into a multi-layered Agentic Orbital Edge Compute architecture.

It simulates a low-power Low Earth Orbit (LEO) satellite performing autonomous perimeter sweeps. Imagery is dynamically acquired via Mapbox (or simulated via SimSat) and analyzed locally on edge hardware (CPU-only) using a fine-tuned LiquidAI/LFM2.5-VL-450M vision language model. The "Orbital Edge" acts as a bandwidth-saver—only escalating critical infrastructure or dense vegetation fire threats to the heavy downstream Ground Station (Dagster + PostGIS) for advanced GIS buffering and risk analysis.

Service Port Map (Local Environment)

Tech Stack & Architecture

GeoFire-Agent/
├── docker-compose.yml           # Bootstraps Postgres, FastAPI, Vision Scout, Dagster, and Vite
├── test_orbital_patrol.py       # Simulates a multi-sector orbital pass hitting the Edge API
├── colab/
│   └── finetune_lfm.ipynb       # Standalone Colab notebook to generate LoRA weights
├── vision_edge/                 # [NEW] Edge CPU visual evaluation microservice
│   ├── app.py                   # Fast inferences on LFM2.5-VL-450M 
│   └── geofire_orbital_weights/ # Mount dir for resulting Colab LoRA adapters
├── backend/                     
│   ├── main.py                  # Hosts the Agentic Patrol logic & API Endpoints
│   └── agent/orchestrator.py    # Multi-node logic triggering Dagster Downlinks
├── dagster_pipeline/            # Heavy raster intersection + buffer jobs
└── frontend/                    # Ground Station GUI

Quick Start

1. Model Fine-Tuning (Mandatory for Vision Edge)

The Local Vision Scout requires an expert adapter.

1. Open colab/finetune_lfm.ipynb in Google Colab (requires free T4 GPU).
2. It will dynamically fetch training images from the Mapbox API to simulate actual satellite tiles, assemble LFM-VL formatted JSONs, and execute a quick LoRA parameter fine-tune.
3. Download the resulting weights from your Google Drive into the local directory ./vision_edge/geofire_orbital_weights/.

2. Bootstrapping the Local Cluster

Once your weights are secured, boot the main orchestration pipeline:

# Clone the repository and navigate inside
git clone https://github.com/hbharani/GeoFire-Agent.git
cd GeoFire-Agent

# Build and start all 5 isolated systems native to Docker Compose
docker-compose up -d --build

3. Initiate the Agentic Patrol

To watch the edge compute logic execute its sweep and conditionally trigger ground downlinks, run the packaged simulation script:

python test_orbital_patrol.py

Note: Depending on your hardware, the vision_edge CPU container can take upwards of ~45s per tile to infer and output its deterministic JSON diagnosis.

4. Review the Ground Station

If an anomaly is escalated by the Scout, track the resulting detailed geo-processing buffers locally at http://localhost:3000 (Dagster UI), and visually observe the flagged vectors at http://localhost:5173 (React Dashboard).

Why CPU-Only Edge Compute?

Hackers naturally scale to the cloud. We restricted the Vision pipeline (vision_edge) explicitly to Python 3.13-slim with strict --index-url cpu PyTorch constraints. This realistically mimics the extreme power and cooling restrictions a small LEO satellite chassis experiences, demonstrating viability for "Space Hackathon" deploy-to-orbit architectures. Bandwidth to Ground hurts; compute on edge saves it.