GeeksforAI

Every disaster report, ranked by need.

An AI reads each report and tells FEMA what to do first.

Email

From: County EOC

Shelter needed for 40 families

Scanned PDF

From: Weather service

Flood warning on River Road

Spreadsheet

From: Hospital

Running on backup power

Priority list

  1. 1Hospital on backup powerHigh
  2. 2Shelter for 40 familiesHigh
  3. 3Close River RoadMedium

Send: power crew, 2 buses, road team

Sample reports, ranked the way our concept works.

What went wrong in the LA fires

When fires hit Los Angeles, help was slow. Not from a lack of data, but from how it moved.

Each county sent its own format.

EmailCounty EOC
WebEOC entryCity EOC
Scanned PDFLocal agency
SpreadsheetShelter
EmailFire department
Scanned PDFHospital

Many departments had to say yes first.

RequestSend more fire crews
  1. Fire departmentWaitingSigned
  2. CountyWaitingSigned
  3. StateWaitingSigned
  4. FederalWaitingSigned

No shared view, so teams could not act as one.

Fire
Police
Red Cross
National Guard
FEMA
The whole picture

Too much data, in too many places, and no one view of it all.

Our answer: one ranked list

The AI reads every report and ranks each task on five things.

  1. How bad it is

    The danger of the event.

  2. How many people

    Everyone in the area.

  3. Key sites

    Hospitals, power stations and roads.

  4. Who is most at risk

    Groups who need extra help.

  5. What is free to send

    Rescue teams, medical teams, supplies.

A person makes the final call on every step.

The concept screens

Four screens we designed for a FEMA operator.

Concept screen: a table of incoming reports by state with priority and status, an AI assessment panel and a dispatch list.
Concept screen. Reports by state, ranked, with the AI’s plan.
Concept screen: a US map with incident markers and a chart of completed evacuations over time.
Concept screen. Incidents on a map, and evacuations over time.
Concept screen: resources deployed, split into search and rescue teams, medical teams and equipment.
Concept screen. Every team and supply that has gone out.
Concept screen: a city map of one incident with live photos and a list of recent requests to approve.
Concept screen. One incident up close, with requests to approve.

How we worked it out

Before any screen, we planned it all on whiteboards.

What it could save

Our team’s estimates, from federal reviews.

  • 0.5 ha day on situation reports, down from 2.5 hFrom GAO report 17-144 and a DHS inspector general report
  • 3 staffreading reports, down from 12From GAO reports 18-472 and 20-236SP
  • $0.7Ma year in labor, down from $1.2MFrom OMB memo M-19-18 and FEMA reviews

How it compares

Next to the tools in use today.

Can itBy handRule basedMap toolStatic dashboardOur AI dashboard
Pull in data liveNoNoYesNoYes
Rank by how bad it isNoYesNoNoYes
Read PDFs, emails and postsNoNoNoNoYes
Suggest what to sendNoNoNoNoYes
Flag reports that disagreeNoNoNoNoYes
Leave the final call to a personYesYesYesYesYes

The pitch

Our DHS DASSH Challenge pitch, with subtitles. I built the AI and ML side.