Surge Sentinel

Built for Track 5 of Build with AI: Code for Communities 2, the cyclone impact and infrastructure vulnerability brief. This is a prototype and a submission, not a system in operational use, and I want to be clear about that up front.

I picked the track because the problem has the shape I keep being drawn to. The data to act earlier already exists. The gap is that nobody can turn it into a specific, local instruction fast enough to matter.

Cyclone response on the Bay of Bengal is mostly post-landfall. The storm arrives, the damage is assessed, and relief follows. The forecast was available days earlier, but a forecast is not an instruction. Knowing a storm is coming does not tell a district officer which road will be impassable, which substation floods, or which shelter to stop sending people to.

So the hours before landfall, which are the only hours when prevention is still possible, get spent on general warnings rather than specific moves.

The useful question is not how strong the storm is. It is which specific things break, and in what order.

That reframing is what makes the problem tractable, because an exposure question can be answered with data that is already public. Terrain decides where surge water goes. Satellite imagery shows what is on that terrain. Forecasts give you the hazard. Put those three together and you can rank infrastructure by risk instead of describing weather.

The prototype fuses Google Earth Engine terrain and satellite layers, SRTM elevation and Sentinel-1 radar, with real-time IMD and ECMWF forecasts to simulate storm surge inundation and rainfall damage pathways.

It then scores the exposure of critical infrastructure against that modelled hazard field: power distribution, arterial evacuation roads, and medical shelters, rendered on a live risk map so the output is a ranked list of places rather than a weather chart.

The last step is the one I cared most about. Gemini's multimodal reasoning reads the hazard field and drafts an advisory in CAP, the standard format emergency systems already consume, addressed to municipal and district disaster management authorities. The point of using the standard is that the output can plug into existing alerting rather than asking anyone to adopt a new tool mid-crisis.

Score infrastructure, not area. A map shaded by flood depth is interesting. A list that says this substation, this stretch of highway, this shelter, in priority order, is actionable. The second one is harder and is the whole value.

Emit CAP rather than a custom format. Less impressive in a demo, far more plausible in reality, because nothing new has to be installed for the advisory to be usable.

Let the model write the advisory, not decide the risk. The hazard modelling and exposure scoring are deterministic and inspectable. Gemini handles the part it is genuinely better at, turning a technical field into clear language for a specific audience under time pressure. I did not want the risk ranking itself to come out of a model nobody could audit.

Anticipatory action is a last mile problem wearing a modelling costume. The hard part was never the inundation simulation. It was deciding what a district officer needs to read at 2am and working backwards from that sentence to the data.

The honest limits: this was built in a hackathon window, the exposure layers are coarser than an operational system would need, and it has not been validated against a real landfall. What it demonstrates is the pipeline and the framing, not accuracy I would ask anyone to act on.