The platform
FireDash was built with the Bilbao Fire Service inside the REACH Incubator, with funding from the Horizon 2020 programme. This is the working version: three screens that answer the three questions a duty officer asks in a shift — what is happening now, what is coming this week, and what has been happening — without asking the service to record anything it does not record already.
One day of operations, replayed minute by minute. Calls arrive, crews commit, the map fills up and the count of available appliances drops. Drag the timeline and the whole screen moves with it.
Live picture · appliances committed · incident log
Expected demand per day, from weather warnings and the city’s own event calendar. The model is on screen — base rate, weather factor, weekday factor, events factor — because a forecast nobody can argue with is a forecast nobody trusts.
Warnings · public events · expected range
The mix a city actually produces: two thirds of the calls are technical assistance, not fires. Time on scene, seasonality, and the hour-by-hour pattern that tells you when a shift is short-handed.
Type mix · duration percentiles · hour and weekday
Under Spanish local government law, fire and rescue is a compulsory service for municipalities above 20,000 inhabitants, and the provincial authority takes it on in the smaller ones that do not run their own (articles 26.1.c and 36.1.c of Law 7/1985). That single threshold produces two very different operations — and the same three questions.
One station or two, a fixed establishment, and demand that is not decided at the town hall. The work is rostering against something that looks random and is not: the hour, the weekday, the storm and whatever is happening in the street that weekend explain a good part of it. And walking into the budget meeting with a figure behind the request instead of an impression.
Ten, fifteen, twenty stations that cannot be compared with each other because each one writes things up its own way. Before anyone can move resources across a territory, the reports have to speak one language — and normalising twenty local vocabularies into one is half the project, and the half nobody wants to do by hand.
Across 4,519 interventions, a single active weather warning changes nothing measurable: 7.74 calls a day against 8.42 on a clear day. Two or more warnings at once is a different matter — 11.59 calls a day, 38% above baseline, over 96 days. That is not a model. It is a fact a duty roster can act on, and it was sitting in data the service already had.
Every call that reaches the service: type, district, appliances, crew, time on scene. This is the spine — everything else is context around it. If you keep them, even in a spreadsheet, you have enough to start.
The public warning feed, by cause and level. It is free, it is already published, and almost nobody crosses it against their own call volume.
The events the council already notifies to the service. A street market and a road race are planned months ahead; the demand they create should not be a surprise on the day.
Assets, hazards and measures follow the NGSI-LD smart data models used across European emergency projects, so the output plugs into what a region already runs instead of becoming one more island.
Not with management. With the person who actually files them, who knows which fields are filled in properly and which ones everybody skips.
What is in there, what is missing, what can already be answered and what cannot. Scoped so the first contract does not have to open a long procurement procedure.
If the history does not support a platform, we say so. A pilot built on records that cannot carry it fails six months in, and by then it is your name on it, not ours.
The reports stay yours. This demo is public and runs on invented records precisely because the real ones are not ours to publish. In a real deployment, where the data sits is a decision for the administration that owns it, and it is the first thing we write down — not the last.
Every record on screen is synthetic. The reports behind this work belong to the Bilbao Fire Service and were shared for research, so we do not republish them. What the demo uses is the shape of that data — hourly profile, type mix, duration percentiles, crew sizes, the weather effect above — measured over 4,519 real interventions between 2019-01-01 and 2020-05-31, and then resampled. It behaves like the real thing because it was fitted to the real thing. The one exception is the map: the footprint is a 150-metre histogram with cells of fewer than three calls suppressed, which is the same kind of aggregate any municipal annual report publishes.
The demo runs on one city. The interesting conversation is what your own records look like once they are in one place, which usually takes a session to find out.
