HurriAid Cover

HurriAid

Why

I’m in Miami. Hurricanes happen. I want a simple, instant view: What’s my risk? What should I do right now? Where’s the nearest open shelter? Also, hurricane myths spread fast. I wanted a quick way to verify rumors without doomscrolling.

What it does

Risk snapshot: SAFE / LOW / MEDIUM / HIGH + a one-line “why.”
Right-sized checklist tailored to your risk.
Nearest open shelter with distance + ETA.
Rumor verifier: labels claims True / False / Misleading with a short note.

How it works (agents)

Watcher: loads advisory → computes distance → asks AI for risk + brief reason.
Parallel:
- Checklist (AI): short, deduped, risk-sized actions.
- Planner: finds nearest open shelter via Haversine distance + ETA.
- Verifier (on demand): fact-checks user-entered rumors.

UI (Streamlit)
  ├─ Watcher → AI risk + why
  └─ Parallel
     ├─ Checklist (AI)
     ├─ Planner (nearest shelter)
     └─ Verifier (rumors)

Run it

# 1) Set your key (example: local dev)
export GOOGLE_GENAI_USE_VERTEXAI=FALSE
export GOOGLE_API_KEY="YOUR_API_KEY"

# 2) Install
pip install -r requirements.txt

# 3) Start
streamlit run app/ui.py

Key math — explained

1) Distance (ZIP → storm center)

Use your ZIP’s latitude/longitude and the storm center’s latitude/longitude.
Compute great-circle distance (follows Earth’s curve, not a straight line).
Result is in kilometers.

2) ETA to the nearest open shelter

Pick a conservative average road speed based on storm intensity:
- TS or Cat 1: ~40 km/h
- Cat 3 or higher: ~30 km/h
- Otherwise: ~45 km/h
Formula: ETA (minutes) = (distance_km / speed_kmh) × 60
This is a safety-first estimate (no live traffic).

3) Risk labels (rule of thumb)

Inputs: advisory radius r, your distance d, storm category, and whether the advisory is active.
SAFE — No active advisory, or you’re well beyond r.
LOW — Advisory is active but you’re comfortably outside the likely impact zone.
MEDIUM — You’re near the edge (≈ within r + 120 km), or inside when it’s TS/Cat 1.
HIGH — You’re inside the advisory, or within ~50 km for Cat 2+.

What I learned

Clean JSON-only contracts keep the UI stable and parsing simple.
Coordinating multiple AI agents (prompts, temps, token limits) matters more than any single prompt.
Small, friendly error messages beat perfect but brittle flows.

Challenges

Getting models to always return strict JSON (no fences/prose).
Robust parsing + retries across agents.
Balancing speed, clarity, and “human” phrasing.

Why it matters

What’s my risk? To classify SAFE / LOW / MEDIUM / HIGH using live advisory data, your ZIP’s distance to the storm center, the advisory radius, and storm category, and include a one-sentence why for transparency.
What should I do now? An AI builds a 24–48h checklist sized to your risk: light reminders for LOW, core supplies for MEDIUM, and full prep + evac steps for HIGH, concise, de-duplicated, and actionable.
Where can I go? We find the nearest open shelter from data/shelters.json, compute distance with haversine, add a storm-aware ETA (slower for CAT3+), and provide a one-click Google Maps route.
Is that rumor real? A verifier labels each claim True / False / Misleading / Caution and adds a ≤30-word note—clear guidance without jargon or scare tactics.