It shows where people get squeezed when a crowd walks around a corner, as a colour map of how crowded each spot is. You describe a corridor bend (its angle, its width, how much of the inner corner is cut off, and how many people walk through per minute) and the map appears at once. Then you change one thing and see whether it helps. The name is Greek for “no way through”.

Anyone who shapes how people move through a space: architects, station and venue planners, event organisers, students, or the simply curious. No crowd science needed: the map and its levels are made to be read at a glance.

Open the demo and pick a scenario, such as a metro transfer or a stadium exit. Each one is a pair: the current situation, A, and a proposed change, B. Move the sliders to adjust either; both maps update as you go, and B's header says how much its peak differs from A's. Hover a map to read the same spot on both.

No. It runs in your web browser, on a laptop or a phone. The first visit downloads about 25 MB, mostly the model; your browser keeps it, so later visits start in a moment. Browsers with WebGPU run it on your graphics card, the others on the processor: a little slower, the same maps.

Each pixel is a density: how many people share one square metre, on average, once the flow has settled. Darker is more crowded. The colours follow Fruin's levels of service, a scale planners have used since 1971, from A, free flow (under 0.31 people per m²), to F, jammed (above 2.15). The number on each map is its peak: the density that only the most crowded 1 % of the corridor goes beyond, so one odd pixel cannot set it.

Because the crowd does not shrink when you redesign the corridor. Aporia keeps the number of people per minute fixed, so a wider corridor spreads the same people over more floor. The difference between A and B then comes from the design, not from fewer people.

In three steps. A research crowd simulator, JuPedSim, tuned on published experiments with real pedestrians, walked virtual crowds through 1,500 different bends. A neural network, CFD-U, then learned to draw the resulting density map directly from the shape of the bend. That network runs in your browser: a simulation takes about 22 seconds, the network about 13 milliseconds, so you can try ideas as fast as you think of them. The network's weights and the 1,500 simulated bends it learned from are public on Hugging Face.

On test bends it never saw during training, from 30 to 120°, its peak is within about 1 % of the simulator's, and it picks the right level of service 99 % of the time. The simulator itself was calibrated and checked against real experiments. The figures are in the numbers above, and the reports below go into detail.

It knows one shape: a single bend in a corridor, everyone walking the same way at an everyday pace. It is validated for angles up to 120°, widths from 1.5 to 3 m and moderate flows; outside that range, the app shows a warning and the error grows. The Disagreement view shows where its three networks disagree: read those areas with care. It models no pushing, panic, counter-flow or evacuation, so it cannot certify a safety plan: use it to compare ideas, then have a real study done.

Free, open-weight (model and dataset on Hugging Face), and no account. Everything is computed on your device: the bends you try are never sent anywhere, and there is no server behind the demo to send them to.

Please get in touch and paste the demo's link: every setting lives in it, so it reproduces your bend exactly.