Inspiration
During an infrastructure crisis, every decision has consequences. Closing a flooded station can protect passengers while preventing others from reaching their destinations. Sending a pump to one location leaves fewer resources available elsewhere.
We built Metro Crisis to explore a question: How can AI assistance help people make better decisions during an infrastructure crisis?
What it does
Metro Crisis is a 3D simulation of subway flooding in a simplified New York City network. Operators deploy pumps and emergency crews, respond to equipment failures, manage station access, and restore service.
Users experience matched scenarios with and without MetroMind assistance. MetroMind prioritizes recommendations and explains them, while the human decides whether to accept or dismiss each one.
The simulator measures passenger hazard exposure, completed journeys, response timing, and spending. A separate replay interface reconstructs recorded runs side by side, connecting decisions to their consequences.
How we built it
We built the browser-based simulator using JavaScript, HTML, CSS, and Three.js.
The simulation tracks water volume across differently sized stations, passenger journeys, train operations, and finite emergency resources. Pumps and crews require travel and setup time. Operational restrictions govern repairs, evacuation, and reopening.
Matching initial conditions and seeded incident schedules support comparisons between conditions. Recorded actions drive the replay, whose final metrics are checked against the original exports.
We used AI coding tools to assist development and debugging, then tested the simulation’s behavior, interface, accounting, and replay consistency.
Challenges we ran into
Our biggest challenge was turning an ambitious concept into a focused infrastructure simulation that could be demonstrated clearly.
We also had to distinguish between assigning equipment and receiving its benefit. For example, a dispatched pump takes time to arrive, and water can keep rising even while it operates.
Making the interface understandable, and comparing assistance fairly without hiding tradeoffs, required substantial iteration.
Accomplishments that we're proud of
We turned our initial concept into a working 3D simulator with different station sizes, finite resources, deployment delays, and operational restrictions. We also built synchronized replay and comparison tools, allowing viewers to see both the decisions and their consequences.
What we learned
Better crisis response involves balancing competing priorities. Across our three recorded matched pairs, assistance was associated with 81.6% lower modeled hazard exposure and $5,500 less spending, but also 410 fewer completed journeys. These exploratory results showed us why human oversight and evaluating multiple outcomes matter.
What's next for Metro Crisis
We want to refine the model with transit experts, collect trials from more participants, and expand the crisis scenarios.We also built an experimental LLM adviser with human approval and action validation. API quota limits prevented reliable live testing during the hackathon, so our demonstrated results use the rule-based MetroMind adviser. Next, we plan to evaluate both approaches.
