OYA AI and the Race Against the Storm

When a storm clears the Caribbean, the bill arrives fast. Dominica lost more than two years of national output to Hurricane Maria in a single night. Across the region, the recovery line in the budget swallows money that should have built schools, clinics, and roads. The Caribbean carries one of the heaviest disaster-recovery burdens on the planet relative to the size of its economies, and the trend is moving the wrong way. Warmer seas feed stronger storms. Rainfall swings harder between flood and drought. The margin for error keeps shrinking.

The world already has good weather science. What it does not have is weather science tuned to an island the size of Saint Lucia. That gap is where OYA AI lives. It sits in the same family of maestro products as Credit Garden, which applies the same physics-informed approach to credit in thin-data markets: build the structure into the model so it holds where data runs out.

The climate-risk bill the Caribbean already pays

Start with the numbers, because they set the stakes. Small island states in the Caribbean routinely lose a share of GDP to climate-driven disasters that would be considered a national emergency anywhere else. A single category-four landfall can cost a country the equivalent of its entire annual output, sometimes more. Insurance covers a fraction of it. The rest comes out of public borrowing, remittances, and household savings that were never meant to absorb a once-in-a-decade shock arriving every few years.

Most of the population lives where the damage lands. Caribbean settlement hugs the coast, the river mouths, and the low-lying flats, because that is where the ports, the tourism, and the farmland are. So a storm does not strike the margins of these economies. It strikes the centre.

Figures above are indicative and illustrative, drawn from public disaster-economics ranges to frame the problem, not audited totals.

Why generic models fail at island resolution

Global weather models simulate the whole atmosphere, and for a hurricane out over open water they will tell you the track days ahead with real confidence. The problem is grid size. A global model carves the world into cells tens of kilometres wide. Many Caribbean islands are smaller than a single cell. To the model, Grenada is barely a pixel.

That coarse grid hides exactly what kills and bankrupts: where the eyewall scrapes a particular coast, how much rain dumps on one watershed versus the next valley over, whether a ridge funnels wind into a town or shelters it. Two villages ten kilometres apart can have completely different storms. A global model cannot see the difference. It hands a government one number for a country that needs forty.

You can run a finer local model on top, but those are hungry. They need dense observation networks, supercomputers, and teams of specialists to keep them honest. Most Caribbean meteorological services do not have that, and they should not need to. The data is sparse precisely where the risk is highest. That is the bind OYA AI is built to break.

What physics-informed AI actually is

Pure machine learning learns patterns from data and nothing else. Give it enough history and it forecasts beautifully. Give it thin, patchy data, which is the Caribbean reality, and it drifts, inventing behaviour the atmosphere would never produce because nothing in its training stopped it.

Physics-informed AI fixes that by building the rules of the atmosphere into the model itself. Conservation of mass and energy, the way moisture condenses, how pressure gradients drive wind, these physical laws sit inside the learning process as hard constraints. The model is not free to guess. It learns from local data where local data exists, and it falls back on physics where the observations run out. The result stays accurate with far less data than a brute-force model needs, and it does not hallucinate a storm that breaks the laws of fluid dynamics.

For islands, this is the answer. You do not need a continent of sensors. You need physics that holds everywhere and learning that sharpens the picture wherever you have a tide gauge, a rain station, or a decade of damage records. OYA AI couples the two so a small meteorological service gets resolution that used to require a national supercomputer programme. This is part of a broader bet maestro is making on the region, the same one behind the AI innovations that will put the Caribbean on the global tech map: build for local conditions instead of importing models trained for somewhere else.

How physics-informed models stay accurate where pure machine learning fails

Thin data breaks ordinary AI in a specific, predictable way, and that failure mode is the whole reason OYA AI exists. A pure machine-learning model is an interpolation engine. Inside the range of its training data it is excellent. Push it to a situation it has never seen, a storm track or a rainfall total outside its history, and it has no anchor. It will produce a confident number that is physically impossible, because nothing in its design knows that mass and energy are conserved. In a region with short, patchy weather records and storms that keep breaking records, the unseen case is the common case. That is exactly when you most need the forecast and exactly when a data-only model is least trustworthy.

Physics-informed models close that gap by carrying their own ground truth. The governing equations of the atmosphere are true whether or not a sensor in Portland or Saint Vincent recorded the last storm. By making those equations part of the loss the model minimises, OYA AI is penalised for any forecast that violates them, even in a place it has no direct data for. Local observations then do a narrower, easier job: they tune the free parameters and correct for terrain and microclimate. The model extrapolates along the physics, not along a guess. In practice that means a usable forecast for a watershed with one rain gauge and forty years of spotty records, the situation across most of the Caribbean, rather than the dense-network situation that mainland models assume.

There is a second benefit that matters for a small country: you can audit it. A black-box model that nails the training set and fails in the field is hard to trust and harder to defend to a cabinet. A model whose forecasts obey known physics can be checked against first principles, which is the kind of accountability we argue for in our work on AI safety from the Caribbean. When the model says the surge will reach a certain elevation, an engineer can reason about whether that is consistent with the wind field and the bathymetry. The physics does not just sharpen accuracy, it makes the output defensible.

A virtual Jamaica you can run the storm through

Underneath everything else sits a virtual model of the island where the laws of physics still apply. Terrain, rivers, coastline, bathymetry, the road network, the power network, land cover and the built environment, assembled into a world a simulation can act inside. Rain falls on it according to the same equations that govern rain. Water runs downhill. Surge stacks against a coast with the shape that coast actually has.

That constraint is what separates this from a chat interface with a weather feed. An AI let loose on an unconstrained model will produce confident nonsense: water pooling on a ridge, a surge that ignores bathymetry, a rainfall total no atmosphere could deliver. Inside a physics-governed world those answers are not available. The simulation can only return outcomes the physical world could actually produce, which is why the scenarios it explores are worth planning around.

Being able to run a scenario changes what a question can be. Instead of asking what the weather will be on Thursday, a parish disaster coordinator can ask what happens to Bull Bay if 600 millimetres falls in 24 hours on ground already saturated by last week. A utility can ask which feeders fail first if a category three crosses the east coast at high tide. A water authority can run a 90-day rainfall deficit against current reservoir levels and see the week the rationing decision arrives. The 2007 or 2024 event can be re-run against today's built environment, with today's roads and today's population, rather than the island as it was when the record was written.

Live in your browser. Terrain, rainfall and water line are a schematic of the modelling loop, not a forecast for any real location.

Swarms of agents that find the risk, then work the fix

A forecast tells you what the atmosphere is going to do. It does not tell you which bridge to close. OYA AI covers that distance with a swarm of AI agents that live inside the virtual island, each one holding a different part of the problem.

One agent reasons about rainfall and runoff. Another about storm surge and the shape of the sea floor near the coast. Another about slope, drainage and where landslides have started before. Others hold the electrical network and which feeders sit under trees on an exposed ridge, the road and bridge network, the health facilities and shelters, the farms and the soil moisture, and the record of what broke the last four times. They run the same event in parallel, thousands of times, with the storm nudged slightly on each pass.

Where the agents converge, you have a risk area. Six of eight agents putting the same square kilometre of Saint Thomas under water in most runs is a finding a parish council can act on. Where they disagree, you get something almost as useful: a place where the data is too thin for anyone to be sure. The swarm flags that gap, requests a satellite pass or a gauge reading, and comes back with a narrower spread. A single model has no way to tell you it is guessing. A swarm shows you its disagreement, and disagreement is where you send the next sensor.

Finding the hotspot is only half the work. The second half is the solution search. Once a risk area is named, the agents propose moves and test each one back inside the physics: close this stretch of the coast road at this hour, de-energise that feeder before the surge arrives, stage the water truck on the upslope side of the bridge that is going to be cut, open the shelter that will still be reachable on Wednesday afternoon. Proposals that fail in simulation are thrown out. What reaches the parish council is a short list that has already survived several thousand simulated storms.

Live in your browser. Agents sweep the model, converge on candidate risk areas, hold a structured consensus, then test fixes. Sites and figures are illustrative.

Flash drought, nowcast from Kingston

Hurricanes get the headlines, but the Caribbean’s other climate threat is quieter and just as expensive: the flash drought. Rainfall stops, temperatures climb, soil moisture collapses in weeks rather than months, and by the time anyone calls it a drought the crop is already failing and the reservoirs are already low. Jamaica has lived this cycle repeatedly. A late, weak rainy season turns into water rationing in Kingston and ruined yields in the parishes.

OYA AI runs flash-drought nowcasting from the same physics-informed core. Instead of waiting for a season to confirm itself, the model tracks soil moisture, evaporation, and rainfall deficit in near real time and projects the trajectory weeks ahead. From a desk in Kingston, a water authority can see a drought forming while there is still time to manage the reservoirs, and a farmer can see it while there is still time to change what goes in the ground. The same model that warns of too much water warns of too little.

From a forecast to a decision

A forecast that sits in a bulletin nobody reads saves no one. The point of extra lead time is the action it buys, and the action looks different for every user. OYA AI is built so each of them can act on the same underlying model.

A parish disaster committee uses the local impact map to time an evacuation, moving people before the road floods rather than during. A relief agency pre-positions water, tarpaulins, and generators on the islands the storm will actually hit, not the whole chain. An insurer prices and pays parametric cover faster because the rainfall and wind footprint is resolved at the level of a community, not a country. A farmer shifts a planting schedule around a drought the model saw coming. A reef monitoring team gets early warning of the marine heat that bleaches coral, and dives to assess before the damage compounds. A family in a vulnerable district simply knows, two days out, whether to stay or go.

None of that works if the forecast speaks only in millibars and probability cones. So OYA AI is designed to put its output in plain language people act on: not “72 percent chance of tropical-storm-force winds” but “expect flooding on the coast road from tonight, move by midday.” The model does the physics. The product does the translation.

What does sharper forecasting do for insurance payouts?

It makes parametric cover pay faster and pay the right people. Traditional disaster insurance sends an adjuster to inspect damage, a process that can take months while families wait. Parametric insurance skips the adjuster: it pays automatically when a measured trigger is hit, say sustained winds above a threshold or rainfall above a level at a defined location. The catch has always been resolution. If the only trigger available is a single national wind reading, the payout fires for the whole island or not at all, and plenty of people who lost everything fall on the wrong side of one coarse number. This is the basis-risk problem, the gap between what the index says happened and what actually happened to you.

OYA AI shrinks that gap because it resolves wind and rainfall at the level of a community rather than a country. A parish that took the eyewall can be paid while a parish that the storm missed is not, from the same event, on the same day. For regional risk pools and the institutions that back them, a credible community-level footprint is the difference between an index people trust and one they suspect. Faster, fairer payouts also change behaviour upstream: a farmer who knows cover will actually pay is more willing to plant, and a lender is more willing to finance. That connects climate forecasting straight back to the credit and growth story we tell in Credit Garden and across our work on AI and the Caribbean economy.

What this changes, and for whom

The point of every hour of extra lead time is the action it buys, and the action is different for each group. Four of them carry most of the value.

There is a fifth effect that does not fit in a grid. OYA AI is not a generic platform bought from a vendor in another hemisphere and pointed at the Caribbean. It is a custom system built in Jamaica, by a Jamaican physicist, for conditions that global models treat as a rounding error. A country that owns its climate intelligence can adapt it when the coastline changes, audit it when a decision is challenged, and keep the skills that maintain it inside its own institutions. That is a different position to be in than renting a forecast.

What does a government need to deploy OYA AI?

Less than most ministries assume. The instinct is that island-scale forecasting requires a national supercomputer and a new directorate, and that instinct is wrong. The physics-informed approach is built precisely so a small country can stand it up with what it already has, plus a few additions that are cheap relative to a single storm’s recovery bill. Three things matter.

First, data. The model needs whatever historical and live records exist: rainfall and tide-gauge series, past storm tracks, and ideally a decade or more of damage and flood records, even if they are uneven. Patchy data is fine, that is the design assumption. What helps most is permission to use it and a path to keep it flowing in near real time.

Second, a handful of well-placed sensors. Not a continental network, a targeted set of rain and stream gauges in the watersheds and on the coasts that drive the most risk. Where a country has gaps, maestro works with it to site a small number of stations where they sharpen the model most. The physics carries most of the load, so a few good sensors go a long way.

Third, partnership rather than replacement. OYA AI is meant to fold into the warnings a national meteorological service already issues, and to sit alongside regional bodies like the Caribbean Institute for Meteorology and Hydrology, not compete with them. The fastest deployments start with the people who already own the forecast and give them a sharper island-scale layer. Governments and agencies can begin that conversation through the maestro get-started page or by reaching the team directly at contact. Training and capacity matter as much as the model, which is why the IMPACT AI Lab exists to bring local builders and officials up the curve.

Where OYA AI stands today

Being straight about the stage matters more than sounding finished. OYA AI is a working prototype with a defined technical concept, a clear set of use cases, a public product explanation and a deployment pathway. It is not yet an operational warning system, and nobody should plan an evacuation on it this hurricane season. What comes next is the unglamorous part that decides whether any of the above holds up.

How communities and governments get access

OYA AI is a maestro product, built so the people closest to the risk can use it without a meteorology degree. A rural high school can pull a local forecast for a class project on climate. A parish council can run scenarios before hurricane season. A national meteorological service can fold OYA’s island-scale layer into the warnings it already issues, sharpening its own forecasts rather than competing with them. Governments and regional agencies can license it at scale; community groups and schools can reach it through maestro’s public access tier.

OYA AI works alongside the experts at the Caribbean Institute for Meteorology and Hydrology and the national services, giving every island, down to the smallest, the resolution that used to belong only to the largest. The storm does not care how big your economy is. Neither should your forecast.

Frequently Asked Questions

What is OYA AI?

OYA AI is maestro’s physics-informed climate product for the Caribbean. It produces island-scale forecasts for hurricanes, flooding, and flash drought, built so communities, schools, families, and governments can plan before an event hits rather than react after it.

What does “physics-informed” AI mean?

It means the model combines learned patterns from data with the physical laws of the atmosphere built in as constraints. Because physics holds even where local observations are sparse, the model stays accurate with far less data than a pure machine-learning system needs, and it will not produce forecasts that break the laws of fluid dynamics.

Why can’t global weather models do this already?

Global models divide the world into grid cells tens of kilometres wide, often larger than a whole Caribbean island. They track storms well at sea but cannot resolve which coast floods or which watershed gets the heaviest rain. OYA AI adds the island-scale detail those models miss.

How much extra warning time does OYA AI give?

Indicatively, OYA AI is built to move locally resolved warning lead time from roughly twelve hours under a generic model to around forty-eight to sixty hours. Those are illustrative figures, not guarantees, and they vary by event and location.

Can it forecast drought as well as storms?

Yes. The same physics-informed core runs flash-drought nowcasting, tracking soil moisture, evaporation, and rainfall deficit in near real time so water authorities and farmers can act weeks before a season fails.

How can my community or government get OYA AI?

Governments, meteorological services, and regional agencies can license OYA AI at scale, while schools and community groups can reach it through maestro’s public access tier. Visit the maestro products page or get started to request access or a briefing.

How would a school or parish council actually use OYA AI before a storm?

Three days out, a parish disaster committee pulls the local impact map and sees which roads and districts are projected to flood and when. Two days out, it times the evacuation order around the road closures rather than after them, and a school decides whether to shelter, close, or release students early. The value is not a single alert but a rolling, place-specific picture people can plan a day around.

What does OYA AI do for flash drought and water utilities?

It nowcasts soil moisture, evaporation, and rainfall deficit in near real time and projects the trajectory weeks ahead. A water authority can begin managing reservoirs and rationing while there is still margin to act, and a farmer can change what goes in the ground before a weak rainy season turns into a failed crop. The same model that warns of too much water warns of too little.

How does OYA AI improve insurance and parametric payouts?

By resolving wind and rainfall at community scale, it shrinks basis risk, the gap between what an insurance index measures and what actually happened to a given household. That lets parametric cover pay the parishes that were hit and not those that were missed, from the same event, in days rather than months. Faster, fairer payouts make farmers more willing to plant and lenders more willing to finance.

How do AI agents fit into a weather model?

They sit on top of it. The physics-informed core says what the atmosphere and the water will do. A swarm of AI agents then runs that event thousands of times inside a virtual model of the island, each agent reasoning over a different domain: rainfall, surge, slope, the electrical network, roads, shelters, soil, and the historical damage record. Where they converge you have a risk area. Where they disagree you have a data gap worth a sensor. Once a hotspot is named, the same agents propose fixes and test each one back inside the physics, so what reaches a decision-maker is a short list that has already survived simulation.

Has this multi-agent method been tested on anything real?

Yes. Section 9, maestro’s Caribbean AI lab, built a nine-agent deliberative system called the Council of Agents and used it to predict Jamaica’s 2025 general election without polling. Each agent reasoned over a different domain and the group was pushed to a structured consensus rather than an average. It called the outcome to within 98 percent accuracy. OYA AI runs the same architecture with physics and terrain in place of constituencies and turnout.

What is a physics-governed virtual world, and why does OYA AI need one?

It is a model of the island where terrain, rivers, coastline, sea floor, roads, power lines and land cover are represented, and where the physical laws still hold inside it. Agents can act in that world but cannot break it, so water runs downhill, surge respects bathymetry, and no simulation returns a rainfall total the atmosphere could not deliver. Without that constraint an AI will produce confident answers that are physically impossible, which is exactly the failure a government cannot afford.

Who built OYA AI?

Adrian Dunkley, a physicist and AI researcher, member of the Climate Studies Group Mona, and founder of StarApple AI and Maestro AI Labs, with 15 years spent building AI systems inside Caribbean institutions. The system is built in Jamaica for Caribbean terrain, coastal exposure and sparse data rather than adapted from a model trained for a continent.

What stage is OYA AI at?

It is a working prototype with a defined technical concept, public use cases and a deployment pathway. It is not an operational warning system yet. The next phase is structured validation: hindcasting against historical Jamaican events using local weather data, damage records and geospatial data, a small set of targeted sensors, and pilot deployments in Jamaica alongside existing warnings rather than in place of them.

What does a government need to deploy OYA AI?

Three things: whatever historical and live data it has, even if patchy; a small set of well-placed rain and stream sensors in the highest-risk watersheds; and a partnership with its national meteorological service rather than a replacement of it. It does not require a supercomputer or a new directorate, which is the point of the physics-informed design.