Wildfire & flood evacuation
Live dispatch for the people who can't drive out.
Alerts tell people to leave. Last Mile finds the households that can't leave on their own, sends the nearest helper who can actually carry them, and re-plans every route as the fire moves.
Simulation with synthetic people · what is simulated
The gap
A warning only helps if you can act on it. Without a car, with a wheelchair or an oxygen line, or on a landline that never got the text, “evacuate now” is not a plan.
In the 2018 Camp Fire, about 40,000 people evacuated and 85 died; NIST's case study examines how notification, traffic and rescues played out. Source: NIST. The CDC notes that evacuation plans and alerts can miss disability-related needs. Source: CDC.
Who this is for
The households an alert alone leaves behind.
People without a car
An evacuation order assumes you can drive away.
Last Mile Matched to a neighbour, volunteer driver, van or shuttle that can reach them before the fire does.
Wheelchair users
Not every ride can take a wheelchair, and boarding takes time.
Last Mile Only wheelchair-capable vans and buses are sent, with boarding time built into the deadline.
Oxygen-dependent residents
Smoke and a long wait are more dangerous; equipment has to come along.
Last Mile Flagged as more urgent on the call, so the dispatcher sees them sooner.
Limited-English households
Alerts in one language reach people late, or not at all.
Last Mile The call happens in the resident’s language and works on a landline. No app, no sign-up.
In ten seconds
Predict. Dispatch. Call.
- 01
Predict
A fire model spreads the front with the wind and gives every household a countdown: how long until the fire reaches that door.
risk clock per home
- 02
Dispatch
The most urgent home that cannot leave alone is matched to the nearest helper who can carry them: a wheelchair needs a wheelchair-capable van. Routes re-plan as the fire moves.
explains every choice
- 03
Call
A Gemini voice agent phones each at-risk home, in its own language, landlines included, and turns the answers into needs on the map.
never promises rescue
How we prove it
Don't trust the demo. Trust the counterfactual.
One good-looking run proves nothing. The results page runs the same scenario under both policies across many seeds and reports the paired difference with its confidence interval, next to the assumptions that drive it. Whatever it says, we show.
Open the Monte Carlo results- 01Same fire
Each seed builds one town, one fire and one set of synthetic people.
- 02Two policies
Run once with alerts only, once with Last Mile. Nothing else changes.
- 03Fair baseline
Last Mile keeps every alerts-only behaviour, family pickups included, and only adds dispatch.
- 04Many seeds
Hundreds of paired runs, reported as a mean difference with a 95% confidence interval.
Built on
- GeminiVoice triage: speech to structured needs
- MongoDB AtlasHouseholds and helpers, $geoNear matching
- Tiger DataRisk and dispatch time series
- VultrRuns the simulation and dispatch engine
- VercelHosts this front end
What is simulated, and what we claim
Last Mile is a working prototype demonstrated in simulation. The full modelling assumptions are listed next to every result.
Read the assumptions- Every resident, helper and household is synthetic. No real person is in this system.
- The fire is a simplified wind-driven model, not an operational fire-behaviour forecast.
- Outcome numbers come only from the Monte Carlo evaluator, run over many paired seeds. Nothing is tuned for the demo.
- If Last Mile did not beat the baseline, the results page would say so.
- The voice agent never promises rescue. If you are in immediate danger, call 911.