Resilience Analytics Lab, LLC
AHI is a calibrated, multi-hazard risk prediction system deployed across the contiguous United States. It produces daily risk probabilities for four hazard types — wildfire, flood, wind, and winter storm — at the county level using a proprietary deep learning architecture trained on 25 years of historical weather, climate, and hazard event data.
The system supports emergency management decision-making at the county and regional level, providing actionable risk tiers, audit-ready outputs, and a structured decision support interface.
- County-level risk predictions across 3,109 CONUS counties
- Five-tier risk classification with operational guidance per hazard
- National overview with state and county drill-down
- Severity-weighted calibration reflecting actual event impact
- 9 regional models serving 48 states + DC
| Source | Coverage |
|---|---|
| NOAA GridMET | Daily weather (temperature, humidity, wind, precipitation, fire weather), 2000–2025 |
| WFIGS Wildland Fire Locations | Historical wildfire records, all CONUS states |
| NOAA Storm Events Database | Flood, wind, winter storm records |
| FEMA Disaster Declarations | County-level disaster records |
This repository contains the deployed dashboard and inference pipeline only. The model architecture, training pipeline, and training data preparation are maintained as proprietary assets.
| File | Description |
|---|---|
app.py |
Streamlit dashboard application |
inference_onnx.py |
Inference engine with per-state calibration |
models/ |
Regional ONNX deployment models |
states/ |
Per-state calibration and reference data |
data/ |
National county reference data |
- Python 3.11
- Streamlit — dashboard framework
- ONNX Runtime — model inference
- Plotly — interactive mapping and visualization
- Pandas / NumPy — data processing
- Deployed on Render
AHI's architecture is grounded in published research on attention mechanisms and topological computation by Joshua D. Curry, available on SSRN:
- Diffusion Attention: Replacing Softmax with Heat Kernel Dynamics
- Heat Kernel Attention: Provable Sparsity via Diffusion Dynamics
- Simplicial Computation: Topology as a Control Variable
MIT License. See LICENSE for details.
The model architecture, training pipeline, and trained weights are proprietary assets of Resilience Analytics Lab, LLC and are not included in this repository.