maestro / Climate / OYA AI
OYA AI is a physics-informed climate intelligence system for Jamaica and the Caribbean. It does not stop at forecasting the weather. It simulates what the weather does to a specific house, road, farm, substation and parish, before the storm arrives.
The prototype below runs in your browser. Pick a storm, run it across the real island, and watch the wind, rainfall, surge, flooding, landslides and damage develop on real terrain.
One road floods while the road above it stays open. One community loses power for three weeks and the next one loses it for three hours. One farmer loses a crop and another has time to harvest. Jamaica already knows this from experience. What the country has not had is a system that can say which road, which community, which farm, and how long before it happens.
Global forecast models track major systems well. They were not built for small islands. A model with 25 km grid spacing sees the whole of Jamaica in a handful of cells, which means it cannot tell the Blue Mountains from the Liguanea plain, or Portland Bight from the north coast. The detail it discards is exactly the detail that decides whether a family needs to leave tonight.
Jamaica needs the step from "a storm is coming" to "this is what it does to your home, your community and your parish, and this is when you should act".
OYA AI builds virtual models of Jamaica and other Caribbean territories, then runs AI agents inside them under physical law. The agents test what-if scenarios before they happen for real: hurricanes, extreme rainfall, flooding, storm surge, drought, landslides and power disruption, and what each of those does to homes, communities, cities and national infrastructure.
Putting a physics constraint on the agents matters. A model free to produce any answer will produce a confident wrong one. A model that has to conserve mass, obey the gradient wind balance and route water downhill can only be wrong in ways you can inspect and correct. That is the difference between a plausible picture of a hurricane and a system a parish council can plan an evacuation with.
What this storm does to this address. Which way the water comes from, how deep, how strong the gusts get on this particular piece of ground, and when to leave.
Which roads become impassable, which shelters sit inside the surge footprint, which schools and clinics lose access, and which communities get cut off first.
Where the wind field crosses the transmission corridor, which feeders sit on slopes likely to fail, and where to stage crews before the rain starts rather than after.
Wind, rainfall and flood resolved at community scale, which narrows the gap between what a parametric index says happened and what people actually lived through.
Rainfall deficit, soil moisture decline and drought stress picked up early enough to change a planting decision or a reservoir release.
How a single event moves through the country hour by hour, so resources are positioned against the sequence of impacts rather than a single headline number.
OYA AI Lite is the modelling layer, running live in a browser tab, with no install and no account. It ships with a terrain model of Jamaica built from 80,729 Copernicus elevation samples. When the page loads it fills every depression in that terrain, routes flow downhill across the whole island, works out how much land drains through each cell and how high each cell sits above the nearest gully, then spreads the 2011 census population across the result.
Then you run a storm at it.
Intensity is bounded by sea surface temperature through the DeMaria and Kaplan maximum potential intensity, cut back by wind shear, and decayed by the Kaplan and DeMaria inland model once the centre crosses the coast. The wind field is the Holland gradient wind profile, with forward motion added on the right of the track where a real hurricane puts it.
Wind is slowed by surface roughness over land, then accelerated again over exposed ridges. Rain follows a radial profile and is then multiplied by the uplift forced when that wind climbs the terrain. The same storm drops a few hundred millimetres on the plain and several times that on the ridge above it.
Storm surge is the inverse barometer rise plus wind setup across the shelf plus wave setup, pushed inland cell by cell and losing height to friction as it goes. Rainfall runoff uses the curve number method, routes down the drainage network, and becomes a water level through the height-above-nearest-drainage method.
Rain-triggered landslides use the infinite-slope factor of safety, with saturation driven by accumulated rainfall and how much land drains through each point. Below 1.0 the slope is expected to go. In Jamaica, slope failure and flooding have killed more people than wind.
Search an address, use your location, or click anywhere on the island. With a Google Maps key connected you get satellite imagery and Street View of the actual building. Either way you get the assessment: gust, rainfall, water depth, slope safety factor, a repair estimate, and the reason this address differs from the next street.
Every equation, every constant, every assumption and every data source is in the methods panel inside the app. The numbers that are tuned rather than taken from a paper are labelled as tuned. You are meant to argue with them.
OYA AI Lite is a prototype of the modelling layer. It is not a forecast, it does not ingest live observations, and it will not tell you what next week's weather will do. Scenarios named after real storms set starting conditions from the public record. The track and intensity you watch are computed forward by the model, not replayed from best-track data.
The value of climate intelligence is measured in hours. Hours between a warning and an evacuation. Hours between a forecast and a crew positioned on the right side of the island. Hours between a rainfall signal and a reservoir release. OYA AI exists to buy those hours and put them where they change an outcome.
Personal risk on a phone changes behaviour in a way a national bulletin does not. A household that can see water reaching its own street evacuates earlier than one told a storm is approaching the island.
Exposure identified before impact is exposure that can be protected. Assets moved, crews staged, stock relocated, equipment raised, and relief positioned where it will be needed rather than where the road still runs.
Community-scale resolution of wind, rain and flood narrows the basis risk in parametric insurance, which is the gap between what the index pays and what the household actually lost.
This is not an imported platform with a Caribbean skin on it. It is a Jamaican system, built for Jamaican terrain, coastal exposure, sparse sensor coverage and real local decision-making, by a Jamaican physicist and AI researcher.
The intensity ceiling comes from the sea, not from a slider:
The wind field is the Holland profile, with the shape parameter recovered from peak wind and pressure drop:
Slope stability is the infinite-slope factor of safety, with saturation driven by the rain the storm has already delivered:
Data sources. Elevation: Copernicus GLO-90 digital elevation model, sampled through the Open-Meteo elevation API. Parish boundaries: geoBoundaries gbOpen ADM1, derived from OpenStreetMap, CC-BY 4.0. Settlements: GeoNames, CC-BY 4.0. Population: Statistical Institute of Jamaica, Population and Housing Census 2011. Satellite imagery and Street View, where connected: Google Maps Platform, using a key the user supplies and which stays in their own browser.
Method sources. Maximum potential intensity: DeMaria and Kaplan (1994). Inland decay: Kaplan and DeMaria (1995). Wind profile: Holland (1980). Radius of maximum wind: Willoughby, Darling and Rahn (2006). Rainfall profile form: Tuleya et al. (2007). Runoff: USDA NRCS curve number method. Flood stage: height above nearest drainage, Nobre et al. (2011). Depression filling: priority flood, Barnes, Lehman and Mulla (2014).
The next stage is structured validation against events Jamaica has already lived through, using historical weather records, damage records, geospatial data and targeted sensor deployment, followed by pilots with the agencies and utilities that would use it. If you run a parish council, a utility, an insurer, a farm, a school or a national agency, and you want to test this against something you remember, we want that conversation.