VANTAGE

📝 Project Overview

Summary of problem, approach, methodology, and findings.

Problem: Disaster response is currently fragmented; during events like Hurricane Beryl, agencies work in silos without a shared data reality, leading to uncoordinated evacuations and preventable deaths.

Approach: VANTAGE creates a "Common Operating Picture" by fusing raw satellite physics with demographic data. We moved beyond simple weather alerts to build a predictive logistics engine.

Methodology: Our 7-stage pipeline ingests live GOES-16 satellite data (Notebook 1), applies radiometric calibration to detect thermal anomalies (Notebook 2), and feeds this into an XGBoost classifier (Notebook 6) to score "Critical Risk" zones. This drives a geospatial optimization engine in the VANTAGE Texas Pilot App that allocates ambulances based on ethical constraints.

Findings: We discovered that purely utilitarian algorithms "abandon" rural areas, necessitating an "Ethical Protocol" selector. We also identified the mathematical "efficiency elbow" for fleet sizing, proving where adding more units hits diminishing returns.

📖 Project Story

Inspiration

2024 was a year of relentless climate chaos—from the Noto earthquake in Japan to the catastrophic floods in southern Brazil and record-breaking heatwaves across the US. These events were a collective wake-up call: our disaster response systems are failing to keep pace with the frequency of modern catastrophes.

For this pilot, we focused specifically on Hurricane Beryl. It was the perfect case study for a broken system: it struck the Caribbean, Mexico, and the USA in rapid succession, yet the response was disjointed. Agencies operated in silos—weather data didn't talk to logistics data, and logistics didn't talk to hospitals. We realized that while borders stop people, they don't stop physics. We wanted to build a system where satellite ground truth—which doesn't lie—could drive a unified, international response intelligence system.

What it does

VANTAGE is a Unified Disaster Response Command Center. It solves five interconnected problems:

• Threat Fusion: Merging satellite telemetry with ground weather data.
• Predictive Risk: Using AI to forecast heat and trauma events before they escalate.
• Ethical Logistics: Allocating ambulances not just by distance, but by fairness and utility.
• Asset Optimization: Preventing duplicate rescues and resource waste.
• Unified Command: Providing a single "glass pane" dashboard for all agencies.

How we built it

We leveraged Hex's multi-notebook architecture to build a modular production pipeline:

• The Physics Engine: We scraped raw NOAA GOES-16 satellite data from AWS S3 using Notebook 1. Using xarray and s3fs in Notebook 2, we performed radiometric calibration to convert digital signals into physical Kelvin temperatures, fusing this with US Census shapefiles in Notebook 3 to calculate population impact.
• The ML Brain: We generated ground truth labels from historical weather data in Notebook 4, engineered time-series features (lag, rolling averages) in Notebook 5, and trained a Champion XGBoost Classifier in Notebook 6. We validated it against historical NWS heat index data to ensure it could predict "High Impact" events with >93% recall.
• The Command Center: We built the frontend app entirely in Notebook 7 (VANTAGE Texas Pilot). We used PyDeck for the 3D geospatial visualization (rendering ambulance density as 3D columns) and built a custom allocation algorithm that users can control via sliders (adjusting fleet scarcity and ethical protocols).

Challenges we ran into

• The "Threshold Trap": Early in our ML training, we accidentally leaked the target variable into our features. We had to rigorously rewrite our labeling logic in Notebook 4 to ensure our model was learning actual patterns, not just memorizing rules.
• Heavy Data vs. Browser: Processing gigabytes of NetCDF satellite files is heavy. We optimized this by building a metadata indexer in Notebook 1 that allows the system to only fetch the exact bytes needed for a specific time window.
• The Ethics of Math: When we built the first optimization algorithm, it strictly maximized "lives saved." We noticed the map turned purple in rural areas—the algorithm was "abandoning" low-density counties to save the city. We had to code an "Ethical Protocol" selector in the final app, allowing the user to choose between "Equity" (fairness) and "Utility" (numbers), making the human cost of the algorithm visible.
• Pipeline Orchestration: We architected VANTAGE as a multi-stage pipeline. A challenge we faced was that Hex projects function independently; updating the upstream "Physics Engine" didn't automatically trigger the downstream "Command Center." We had to manually chain outputs, identifying a need for a native "Cascade Run" feature for dependent projects.

Accomplishments that we're proud of

• Disaster Agnostic Architecture: While we piloted with Hurricane Beryl, 90% of our pipeline is event-agnostic. With the correct data, the system can be adapted to simulate wildfires, floods, or typhoons with minimal pipeline changes.
• Accuracy: Achieving >93% Recall on our XGBoost model, meaning our system rarely misses a critical heat event.
• Speed: Successfully fusing 1TB+ of satellite potential into a lightweight, responsive dashboard that loads in seconds.
• Ethical AI: We didn't just build a logistics tool; we built a moral compass. Our "Ethical Protocol" selector explicitly visualizes the trade-offs in disaster response, ensuring rural communities aren't invisible to the algorithm.
• 3D Visualization: Implementing a seamless PyDeck 3D map entirely within the Hex environment to visualize resource density intuitively.

What we learned

We learned that disaster response is a logistics problem, not just a weather problem. Accurate weather predictions are useless if you can't optimize the allocation of the ambulance that responds to them. VANTAGE proves that when you break data silos, you don't just see the storm clearer—you see the path to saving lives.

What's next for VANTAGE

• International Expansion: Scaling the Texas Pilot to Jamaica and Mexico to test cross-border coordination capabilities.
• Live Validation: Integrating ground-level social media scraping to visually validate the "Thermal Anomalies" detected by our satellite physics engine.
• Automated Pipelines: Working to implement a true "Cascade Run" feature to make the entire ingestion-to-deployment pipeline fully autonomous without human intervention.

Built With

• geopandas
• hex
• noaa-goes-16
• open-meteo-api
• pandas
• plotly
• pydeck
• python
• scikit-learn
• xarray
• xgboost