72h
Target hurricane lead time vs 6 hours current
$64B
Caribbean climate losses 2000-2023
2.9%
GDP lost per major storm event
$130B
Annual reinsurer losses from underpricing

The warnings exist and the models exist. They are too slow, too coarse, and they were not built from Caribbean Sea data. Every operational hurricane forecast model in use today was trained primarily on Atlantic and Gulf of Mexico conditions. The Caribbean Sea has specific thermal dynamics, shallow bathymetry, and orographic interactions with the island chain that behave differently from the open Atlantic. Existing models treat these as secondary inputs. OYA AI treats them as the primary signal.

The result: 6 to 8 hours of usable warning today, against a target of 72. That gap changes the kind of response a country can mount. 6 hours means shelter in place. 72 hours means structured evacuation, supply pre-positioning, infrastructure protection, and economic activity planning. The difference in casualties and economic damage between those two scenarios is the entire case for OYA AI.

Named for Oya

OYA AI is named for the Yoruba goddess of storms, transformation, and change. She stands at the crossroads between life and death, whose domain is precisely the kind of sudden, devastating, clarifying force that a Caribbean hurricane represents. The name is intentional. The relationship between the storm prediction system and the communities it serves is not neutral or purely commercial. It is consequential in a way that the name acknowledges.

How It Works

OYA AI is a physics-informed AI system. The model encodes actual atmospheric and oceanic dynamics rather than fitting statistical patterns to historical storm tracks. That distinction matters for a specific reason: pure pattern-matching models fail when the distribution shifts. As Caribbean sea surface temperatures change decade by decade, a statistical model calibrated on historical storm behaviour drifts from reality. A physics-informed model remains accurate because the underlying physical relationships do not change when the inputs do.

The system runs on continuous data ingestion: Caribbean Sea surface temperature at sub-kilometre resolution, atmospheric pressure gradients, wind shear measurements, ocean heat content, and satellite imagery updated on six-hour cycles. The nowcast output updates every cycle and is available to API subscribers in real time.

"The social mission and the commercial case are the same product. The data that saves lives is the same data that reinsurers will pay to price Caribbean risk accurately. One pays for the other by design."

The innovation

OYA AI: physics-informed climate intelligence for Jamaica and the Caribbean

OYA AI builds virtual models of Jamaica and other Caribbean territories, then sends AI agents into those physics-governed worlds to test what a storm, a flood or a dry season would do to a named road, community, feeder line or watershed before it happens for real.

Built by
Adrian Dunkley, physicist and AI researcher, member of the Climate Studies Group Mona, founder of StarApple AI and Maestro AI Labs
Behind it
15 years building AI systems inside Caribbean institutions rather than importing them
Method
Climate physics, geospatial intelligence, machine learning, virtual modelling and agent-based simulation in one loop
What it forecasts
Impact, not only weather: who floods, which road closes, which feeder fails, which watershed takes the heaviest rain
Designed for
Caribbean terrain, coastal exposure, sparse sensors, uneven data and decisions made at parish scale
Stage
Working prototype with defined use cases and a deployment pathway. Structured validation and Jamaican pilots come next.

The virtual island

Underneath the forecast 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 wired to a weather feed. An AI let loose on an unconstrained model produces 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, which is why the scenarios the system explores are worth planning and pricing around.

Being able to run a scenario changes what a question can be. 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 reinsurer can re-run the 2007 event against today's built environment and today's exposure rather than the island as it was when the record was written. That last capability is the one that turns a warning system into a risk-pricing asset.

OYA AI / virtual island, physics-governed Simulating
Scenario 01 / category three landfall, east coast Water line +0.00 m

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

Agent swarms that find the risk and work the fix

A forecast tells you what the atmosphere is going to do. It does not tell you which bridge to close or which portfolio is exposed. 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, 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 and an underwriter can price. Where they disagree, you get something almost as valuable: a place where the data is too thin for anyone to be sure. The swarm flags that gap and requests a satellite pass or a gauge reading. A single model has no way to tell you it is guessing. A swarm shows you its disagreement, and disagreement is where the next sensor goes.

Finding the hotspot is 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 discarded. What reaches a customer is a short list that has already survived several thousand simulated storms.

OYA AI / agent swarm, risk resolution Deliberating
Phase 01 / agents sweep the virtual island 150 agents · 5 risk areas · agreement 0%

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.

Agent 01

Rainfall and runoff

Where the water lands, how fast it moves, and which watershed takes the load.

Agent 02

Surge and coast

Wind stress, tide state and the shape of the sea floor that decides how far the water reaches inland.

Agent 03

Slope and drainage

Saturation, gradient and the hillsides that have moved before.

Agent 04

Electrical network

Feeders, substations and the spans exposed enough to fail first.

Agent 05

Roads and access

Which routes get cut, in what order, and what that isolates.

Agent 06

Shelters and health

Capacity, reachability and the facilities that sit inside the projected impact.

Agent 07

Soil and agriculture

Soil moisture, crop stage and the fields that lose a season rather than a week.

Agent 08

Historical damage

What actually broke last time, and how far the model drifted from it.

Where the method comes from

98% accuracy, Jamaica general election, 2025

The swarm is not an idea borrowed from a paper. Section 9, Maestro’s Caribbean AI lab, built a nine-agent deliberative system called the Council of Agents and pointed it at Jamaica’s 2025 general election. It ran no polls. Each agent reasoned over a different domain: constituency-level voting history, demographic shift, media and voter sentiment, turnout patterns. They disagreed, argued, and were pushed to a structured consensus instead of an average.

The Council called the outcome to within 98 percent accuracy. OYA AI runs that same architecture on a different substrate. Swap constituencies for watersheds and turnout for rainfall and the machinery is unchanged: many narrow specialists, a governed world to argue inside, and a consensus you can read back line by line to see which agent said what and why. For a reinsurer pricing a treaty or a minister ordering an evacuation, that audit trail is the difference between a number and a defensible number.

Three Buyer Markets

Reinsurers lose $90 to 130 billion annually from Caribbean and Latin American climate risk they cannot accurately price. Most of that figure is pricing error from insufficient data rather than loss from the storms themselves. OYA AI's climate data platform provides the actuarial input that makes accurate pricing possible. A single large reinsurance client reducing its Caribbean loss ratio by 5% through better pricing data represents $225 million in annual value.

Governments in the Caribbean and LATAM spend approximately $2.4 billion annually on disaster preparedness and emergency management. A 12x improvement in warning lead time has a direct, measurable return on investment through reduced emergency response costs, reduced structural damage to public infrastructure, and fewer casualties per storm event.

Development banks including the Inter-American Development Bank, Caribbean Development Bank, and World Bank allocate $8 to 12 billion annually to climate resilience investments in the Caribbean and LATAM region. Accurate climate intelligence is a prerequisite for effective investment allocation. OYA AI provides that signal where it does not currently exist.

Addressable Market

The global catastrophe risk data and analytics market is $12.7 billion, projected to reach $22.1 billion by 2030. Caribbean and LATAM coverage is currently the largest single data gap in that market. OYA AI occupies that gap directly.

Government emergency management technology spending in the Caribbean exceeds $400 million annually. LATAM adds $1.8 billion. The reinsurance premium market for Caribbean risk exceeds $4.5 billion annually. Each of these is a distinct revenue channel that the same underlying product addresses.

Revenue Model

OYA AI generates revenue through government API contracts, reinsurance data licensing, development bank data partnerships, and emergency management technology contracts. The data licensing model scales with the number of jurisdictions covered and the frequency of data delivery. Government contracts provide baseline recurring revenue. Reinsurance partnerships generate premium-linked fees that scale with portfolio size.

Where the system 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. The figures on this page are targets and market estimates, not delivered performance. What comes next is the unglamorous work that decides whether any of it holds.

Now

Prototype and product definition

Physics-informed core, virtual island modelling, agent swarm architecture, and the use cases each one serves.

Next

Hindcast against real events

Run the system backwards over historical Caribbean storms, floods and dry seasons, then score it against what happened: damage records, flood extents, outage logs and crop losses. A model that cannot reproduce 2007 has no business forecasting 2027, and no reinsurer will price off it.

Then

Targeted sensing and data partnership

A small number of rain and stream gauges sited where they sharpen the model most, plus geospatial and observational data from national meteorological services and the agencies holding the damage record.

After

Pilot deployment in Jamaica

One or two parishes, one utility, one insurer, running alongside existing warnings rather than in place of them, with every call scored against the outcome. Regional expansion follows the evidence.