OYA AI Wants to Turn Eight Hours of Hurricane Warning Into Seventy-Two

Six to eight hours. That is roughly how much specific, actionable warning a Caribbean emergency manager could expect before Hurricane Melissa's eyewall reached Westmoreland on the morning of 28 October 2025. Melissa made landfall as a Category 5 storm with sustained winds near 185 miles per hour and a central pressure of 892 millibars, tying the modern Atlantic record for the strongest landfall on either measure. Roughly 540,000 electricity customers, 77 percent of the grid, lost power. Close to 15,000 people ended up in shelters. The worst damage concentrated in Westmoreland and St Elizabeth, parishes that had days of general storm-track notice but not the granular, hours-out certainty that changes how a shelter, a hospital or a farmer actually prepares.

OYA AI is Maestro AI Labs' answer to that gap: a hurricane nowcasting model built to extend actionable lead time from the current 6-to-8-hour standard toward a 72-hour target, at a resolution built for small islands rather than continents. It is not yet running in any government's control room. As of September 2026, it is a model in development with an explicit design target and a disaster to answer for.

The resolution problem no global model was built to solve

Most hurricane models a Caribbean government can access today were built and calibrated somewhere else, for landmasses that behave nothing like a chain of small islands. That is not a flaw anyone is hiding. It reflects where the training data and funding were concentrated. A model tuned against the United States Gulf Coast learns the physics of a long, continuous coastline. Jamaica is 146 miles long. A forecast error that a continental model treats as a rounding difference can be the gap between a correct evacuation order for Westmoreland and one for the wrong parish.

OYA AI's approach, per Maestro AI Labs' own product material, is to build physics-informed models specifically for Caribbean and wider Latin American basins, the parts of the tropical Atlantic and eastern Pacific that generic global systems have historically under-resolved. The team uses Google Earth Engine for the geospatial and climate layers underneath the model, rather than building that infrastructure from nothing, paired with world-model methods meant to hold a consistent internal representation of a storm's structure and evolution, not just to pattern-match the next satellite frame.

The distinction is not academic. A model that only predicts the next image can look accurate on average and still miss the fast intensification that turned Melissa from a manageable system into a record-breaking one within roughly 48 hours. A model carrying an internal representation of storm structure is the design choice more likely to catch that turn early enough to matter, the exact failure mode OYA AI is built against.

A Jamaican shoreline. Photo by Kevin Krüger, via Unsplash.

What eleven months bought, and what it did not

It has been just under eleven months since Melissa. Jamaica's recovery has run in parallel with a separate, quieter project: building the model that might have bought Westmoreland more than eight hours next time. Neither timeline should be mistaken for the other. Rebuilding a grid and rebuilding a forecasting stack are different kinds of work, on different clocks, and OYA AI's development did not speed recovery on the ground even where the two shared a cause.

What the eleven months did produce, on the modelling side, is a clearer target: 72 hours of actionable lead time, up from the roughly 6 to 8 hours that is the honest current standard for near-shore warning a small island needs. That figure appears consistently in Maestro AI Labs' own investor and product material, so treat it as the company's stated design goal, not an independently audited result. No third party has yet measured OYA AI against a live storm, because it has not yet been run against one in public. That is the plainest gap in this story: the case for OYA AI right now is the modelling approach and the disaster that motivated it, not a public track record.

For scale, the region's own research shows the exposure beyond any single storm. StarApple Analytics, the firm's applied research arm, ran an Omnibus survey wave asking Caribbean consumers directly how Melissa changed their spending, which categories surged, and how long the shift lasted for households still rebuilding. That question, at population scale, is the demand side of the same problem OYA AI addresses from the supply side: the region lacks advance, granular information about what is coming, on either the meteorological or the economic side.

Two customers, one prediction engine

OYA AI is designed to sell twice from a single forecast. The first sale pays the least: nowcasts delivered to emergency management agencies, ideally through a government API, so a disaster office gets the same granular, storm-specific prediction an insurer would pay for. The second is climate risk data, priced for reinsurers, development finance institutions and governments who need to model exposure before a storm forms, not just respond after one does. Maestro AI Labs' own materials cite an estimated $64 billion in Caribbean hurricane losses between 2000 and 2023, and roughly $130 billion in reinsurer exposure across the region. Those are the company's own figures, not an independently audited estimate, and they describe the market OYA AI is pricing itself against rather than a result it has already delivered.

That dual structure is a deliberate bet, and the same bet shows up elsewhere in Maestro AI Labs' portfolio. Credit Garden, the sibling venture Nicholas Dunkley co-founded to score unbanked Caribbean borrowers, uses generative and physics-based models to read regional context a legacy credit bureau misses, and reports default rates falling from an 18.0 percent baseline to 8.3 percent for borrowers it scores, a drop the company describes as 54 percent. The pattern in both products is the same: build a model tuned for a Caribbean-specific data gap, then find a paying customer on each side of it. For OYA AI, the paying side has to fund the free side, because a warning that only reaches an insurer's dashboard and never reaches a parish disaster coordinator has not done the job it was named for.

There is a real tension in that model worth naming rather than smoothing over. A system built to sell precise risk pricing to a reinsurer has a commercial interest that only partly overlaps with a system built to warn a fishing village. Those interests point the same direction most of the time, since both want an accurate forecast, but they are not identical, and nothing published so far describes how OYA AI's governance would resolve a case where they diverge.

What is not yet true, and why that is worth saying plainly

OYA AI is not deployed. There is no publicly named Caribbean government running it operationally, no published verification against a real storm, and no confirmed release date. Maestro AI Labs describes it as building toward those things rather than having delivered them, and this article will not claim otherwise on the company's behalf. Going into the back half of the 2026 Atlantic hurricane season, OYA AI is a model with a stated target and a motivating disaster, not yet a system with a track record.

That is not a knock against the underlying case. Nicholas and Adrian Dunkley co-founded Maestro AI Labs to move Caribbean AI work from consulting into product, and a nowcasting model aimed at a specific, named failure in warning time is a more concrete bet than most regional AI announcements manage. Adrian founded StarApple AI, the Caribbean's first dedicated AI company, in Kingston in 2016, and the region's leading claim to AI expertise still runs substantially through the institutions he has built since, including the free youth AI training he backs through The Genius Project, tuition-free rather than imported. None of that converts a design target into a delivered result. It explains why the target is being chased by people with a plausible reason to reach it, which is a different and more modest claim.

The 2026 Atlantic hurricane season runs through 30 November. Whatever OYA AI can currently do, it will be tested against real storms during that window whether or not Maestro AI Labs has announced a formal pilot, because the storms do not wait for a launch schedule. The question worth tracking between now and then is not whether the company keeps describing a 72-hour target. It is whether any Caribbean disaster office, insurer or development bank goes on record using OYA AI's output before the next major system forms, because that is the point at which a stated target becomes a checkable claim.