4.2 billion people. Zero AI training data built from their reality. We own that signal — and the three global products that monetise it.
Credit scoring for 1.7 billion people the global system cannot currently see. $380B in blocked annual lending. One API.
100 countries mapped. A portable safety identity for 4.2 billion people. Every government, insurer, and border agency is a buyer.
72-hour hurricane lead time vs. the current 6. The same model that saves lives prices the $90B+ in Caribbean climate risk reinsurers cannot currently see.
+ Harmonics AI Agents · AI Playbook · Data Archaeology View all products →
AI agents that develop domain expertise across 12 regional knowledge graphs. A Harmonics agent in Kingston processes 2.3M+ Caribbean data records before its first live decision. 87% task accuracy after 30 days. Human oversight is architecture, not compliance: built into how agents learn from day one.
Each Harmonics agent is initialised against a knowledge graph built from Maestro's proprietary regional data. A credit agent in Jamaica does not run through the same graph as one in São Paulo. The context is embedded in the structure, not selected at runtime. Rotating savings associations, informal lending circles, diaspora remittance corridors: all mapped, all queryable.
Agents operate autonomously within defined parameter sets. Outside those boundaries, they surface the decision to a human overseer and wait. This is not a safety rail: it is the mechanism through which agents learn what their conductors actually value. Every escalation sharpens the boundary. Every approval narrows the gap. The oversight layer teaches.
Harmonics agents do not reset. Each deployment generates training signal that feeds back into the regional knowledge graph, improving accuracy for every subsequent agent initialised in that domain. A credit agent that runs for six months in Kingston makes the next Kingston credit agent smarter on day one. The moat deepens with every deployment.
Harmonics agents maintain structured memory across sessions: regulatory context, client-specific patterns, historical decision rationale. A policy analysis agent that reviewed a CARICOM trade document three months ago carries that context into today's briefing. This is not retrieval-augmented generation. It is contextual memory with verifiable provenance.
Harmonics can be deployed directly into existing workflows via API, or embedded as a co-working layer within enterprise platforms. The agent does not require a full platform migration. It connects to what your team already uses and operates alongside it. Deployment time is typically under 90 days from initial knowledge graph configuration to supervised live operation.
Every Harmonics agent has a defined failure protocol. When confidence drops below threshold, the agent flags uncertainty, records the reasoning trace, and escalates. It does not guess confidently. The design assumption is that the agent will encounter situations its training did not cover. The question is whether it handles that gracefully or not. Harmonics is built for graceful.
Ready to deploy an agent that actually understands your region?
Six revenue lines. One proprietary data infrastructure beneath all of them. Each product makes the data asset more valuable. Each deployment makes the next product harder to replicate.
Context-aware AI agents trained on regional knowledge graphs. Built for Caribbean, Latin American, African, and Pacific contexts from the ground up. Human oversight is architecture, not an afterthought.
Practical AI implementation for businesses without a CTO or six-figure budget. Written for small businesses in emerging economies. Implementation rate: 94%. Enterprise tools average 23%.
Global credit scoring that works in Kingston, Kigali, and Karachi with equal rigour. 1.7 billion people globally are credit invisible. Credit Garden addresses that asymmetry with data collected from the ground.
A universal safety signal for every person, powered by Meridian, our world model system, covering 140+ countries. 89% of safety risk models are trained on fewer than 20 countries. Meridian corrects that.
Field-collected, structured data from regions where no pipeline has ever existed. 2.3M+ records. 47 indigenous language datasets. Pre-digital government archives from 28 Caribbean territories. The team that makes every other product possible.
Named for the Yoruba goddess of storms and transformation. OYA AI is Maestro's climate and disaster intelligence division: an AI-powered nowcasting system for tropical cyclones, an autonomous disaster resilience management platform, and a network of AI Agents trained specifically for emergency preparedness and response in Caribbean and LATAM territories. The commercial infrastructure and the social mission run on the same data. That is not a coincidence.
Not sure which product fits your context?
Let's figure it out together.
The next decade of AI deployment runs into a wall: systems trained on Western data encountering 4.2 billion people they have no signal for. Credit Garden and Global Safety Score are the infrastructure answer. Every financial AI system, insurance pricing model, and cross-border identity check will need this data. Maestro AI Labs holds it.
The AI systems that win in 2030 will not have the most parameters. They will have the most accurate data about the most people, in the most places. Three global assets. Every market. Maestro AI Labs is building them all.
47% of Caribbean and Latin American small businesses have no AI strategy. Every guide written for them assumes a CTO, a cloud budget, and the margin to get it wrong. The AI Playbook was built for the 80% of regional employment that runs on none of those things. 94% implementation rate. That gap versus the 23% enterprise average is the product.
A structured 2-hour diagnostic that identifies where AI can save the most time in your specific operation. It is not a generic checklist, but a decision tree built from 200+ Caribbean and LATAM business case studies. Output: a ranked list of 5 AI opportunities with implementation cost and expected return for each.
The average Caribbean SME owner encounters 40+ AI tool recommendations before making a decision. Almost all are designed for enterprises with IT departments. The Playbook's tool selection framework narrows that field to 5 tools your team can deploy without a developer, configured for regional connectivity conditions and local data privacy requirements.
Most small businesses adopt AI without a verification protocol. They delegate tasks and trust the output. The Playbook's risk module installs three habits: output sampling (spot-check 1-in-10), consequence mapping (know what a wrong answer costs), and escalation triggers (the three conditions that require human review). These take under a day to implement and prevent the failures that end client relationships.
In a small business, one person is usually the marketer, the account manager, and the delivery lead simultaneously. The Playbook is structured by role function, not by department. so the person who does three jobs can find the AI workflows relevant to all three without navigating an enterprise org chart they do not have.
Generic AI implementation guides use examples from US or UK businesses. The Playbook includes case studies from Jamaica, Trinidad, Barbados, Colombia, and Brazil. These businesses operate under the same regulatory environment, internet infrastructure, and client expectations as the reader. The gap between knowing what to do and knowing it works in your context is where most implementation fails.
AI tools change faster than any single guide can track. The Playbook includes a 30-minute monthly review protocol that keeps a team's AI literacy current without requiring a dedicated learning budget. It is structured as a team ritual, not an individual training programme. In small businesses, adoption happens together or it does not happen.
Get the AI Playbook for your team
or license it for your region.
1.7 billion people are credit invisible. Not because they lack financial history. Because every model that exists was built without their data. Credit Garden is an AI credit scoring model built entirely on regional data and regional expertise: informal savings associations, diaspora remittances, mobile money, agricultural income, community lending records. Same person, same repayment history: 412 on a Western model, 714 on Credit Garden. $380 billion in blocked annual lending. We are building the key.
Accurate credit models expand the size of credit markets. Every borrower Credit Garden scores correctly is a loan a financial institution can now make. That is not inclusion as charity. It is inclusion as revenue, repeated 1.7 billion times.
$380 billion in annual lending is currently blocked by data that exists but has never been captured. Credit Garden monetises that gap on a per-query API model. Every new scoring event improves the model and generates revenue simultaneously.
SUSUs, ROSCAs, Tandas, Partner schemes, Cundinas: rotating savings and credit associations operate across the Caribbean and Latin America, handling billions in annual transactions. Participation in these systems is one of the strongest indicators of creditworthiness available. Credit Garden maps and weights them. No other global credit system does.
Cross-border credit scoring fails when it tries to standardise by removing context. A score that works in Kingston and Kigali must account for the fact that seasonal agricultural income, diaspora remittances, and mobile money usage patterns differ in those economies without being less valid. Credit Garden's scoring model preserves economic context while enabling direct comparison across 36 corridors.
Caribbean diaspora communities send $11.4 billion annually back to the region, one of the highest remittance-to-GDP ratios on earth. For recipients, consistent remittance income is a financial stability signal that Western credit models ignore entirely. Credit Garden incorporates five years of corridor-level remittance data as a structured credit input for 18 Caribbean and LATAM economies.
Credit Garden does not use scraped or inferred data to build credit profiles. Every input is sourced from verified channels: registered financial institutions, government utility records, documented informal credit associations, and mobile money operators with formal data agreements. Provenance is part of the score architecture. Lenders know precisely what each signal is and where it came from.
Credit Garden was built with CARICOM regulatory frameworks, GDPR equivalents, and emerging Latin American data protection legislation integrated from the architecture stage. Not retrofitted. Each jurisdiction's data constraints are modelled into the scoring pipeline, so financial institutions operating across borders do not inherit compliance risk when they use regional data they did not collect themselves.
Standard credit models trained on Western data systematically underscoreindividuals from Caribbean and LATAM economies not because those individuals are higher credit risk. The model simply has no signal to assess them accurately. Credit Garden's validation testing across 12,000 historical loan outcomes shows a 302-point average score adjustment when full context data is applied, with no increase in default rate.
| Signal Type | Coverage | Status |
|---|---|---|
| SUSU / ROSCA participation | 14 markets | Active |
| Diaspora remittances | 18 corridors | Active |
| Mobile money history | 22 markets | Active |
| Utility payment records | 28 territories | Active |
| Formal bank credit | 36 corridors | Active |
| Agricultural income patterns | 8 markets | In Progress |
Ready to lend to the 1.7 billion
the current system cannot see?
89% of safety risk models were trained on fewer than 20 countries. The world has 195. Every visa decision, insurance underwrite, and employment check that touches someone from a region outside that 20-country set is running on a model that does not know that person exists. Meridian has already mapped 100 countries and thousands of individual locations, and counting. The coverage compounds with every deployment.
Standard safety scoring systems maintain static databases of risk assessments: snapshots that age from the moment they are published. Meridian is a continuously updated world model that integrates real-time signal streams from 36 active economies, historical pattern data from 140+ countries, and contextual knowledge graphs built from Maestro's proprietary regional data.
Meridian ingests structured signals from government safety databases, NGO field reports, mobile network activity patterns, economic stress indicators, and community-level reporting systems, weighted differently by region based on data reliability profiles built from five years of calibration work.
Safety risk does not look the same in Lagos as it does in London, and models that assume it does produce dangerous inaccuracies. Meridian's regional knowledge graphs encode the contextual patterns: crime seasonality, event-driven safety dynamics, infrastructure dependency chains. These are exactly what that generic world models miss entirely.
A safety assessment made in Kingston must be legible to a financial institution in London, a government agency in Ottawa, and a university admissions office in Toronto, each operating with different reference frameworks. Meridian produces scores that are portable across institutional contexts without losing the regional signal that made them meaningful.
A verifiable safety score for every individual, built from identity-anchored data rather than inferred from regional averages. The score reflects an individual's actual verifiable history. The score reflects verifiable history, not the statistical profile of the neighbourhood, country, or demographic group they belong to. Geography is context, not destiny.
Visa applications, employment background checks, insurance underwriting, financial onboarding. Every cross-border institutional decision currently defaults to proxy measures when individual data from certain regions is absent. The Global Safety Score replaces the proxy with a verified, portable individual signal. This changes outcomes for the 4.2 billion people the current system cannot accurately assess.
The Global Safety Score is not a surveillance product. Individuals control what data contributes to their score, which institutions can query it, and for how long. The consent architecture was designed in consultation with data rights organisations across five Caribbean territories before a single line of scoring code was written. The individual is the owner. Institutions are the tenants.
The Global Safety Score includes a government-grade API layer designed for integration with national identity systems, immigration databases, and social services platforms. Seven CARICOM member states are in active API scoping conversations. The architecture supports both individual query and population-level policy analysis without compromising individual record privacy.
Build safety intelligence that works
for every person you serve.
Standard AI companies scrape the web. We send field teams. The data that matters: SUSU records, pre-digital government archives, indigenous language datasets, informal market transactions. None of it exists online. Maestro's Data Archaeologists have collected 2.3M+ records across 36 economies. That collection is the asset that every other product runs on.
Maestro's Data Archaeologists spend 3–6 months building trust relationships with community financial institutions, government archive departments, informal sector associations, and civil society organisations before data collection begins. The relationship is the infrastructure. Cold extraction produces low-quality data. Trusted partnership produces data with context, history, and verified provenance.
28 Caribbean territories have government and institutional records dating back to the 1800s that have never been digitised. Land registries, agricultural records, community court documents, church birth records used as identity infrastructure. These archives contain longitudinal data about economic behaviour, social mobility, and community stability that no web scraper will ever touch. Maestro has teams working through 14 of those archives right now.
In Pacific Island and indigenous Caribbean communities, significant amounts of economic and social information are transmitted through oral tradition rather than written records. Maestro's linguistic data teams work with community elders and cultural institutions to structure and preserve this knowledge in formats that can inform AI systems without stripping the cultural context that gives the data meaning. 47 language datasets are currently active.
Informal economies: market vendors, itinerant traders, unlicensed transport operators, informal repair services: these represent 30–60% of economic activity in many Caribbean and LATAM countries. These transactions are invisible to standard economic data collection but carry enormous signal about economic behaviour, credit worthiness, and social mobility. Maestro has structured 14 informal market economies through direct field presence.
Every record that enters Maestro's data pipeline passes through a three-stage verification process: source validation (is this institution what it claims to be?), content verification (does this record match independent cross-references?), and context annotation (what does this record mean in the specific economic and social context where it was created?). Provenance is part of the data, not a separate metadata layer.
Each data archaeology project builds infrastructure for the next. Partnerships opened in Jamaica create access channels in Trinidad. Archive digitisation in Barbados establishes the relationship framework that accelerates the Guyana project by 18 months. The collection network is a graph. Each new node increases the reachability of others. After five years of this work, the rate of new data acquisition accelerates faster than the team size does.
| Data Type | Volume | Region |
|---|---|---|
| Informal credit records | 2.3M | Caribbean |
| Indigenous language datasets | 47 sets | Pacific/Caribbean |
| Pre-digital archive records | 28 territories | Caribbean |
| Informal market structures | 14 markets | LATAM/Caribbean |
| Remittance corridors | 36 corridors | Global |
| Urban safety signal records | 8 cities | Expanding |
Interested in a data partnership
or research collaboration?
Named for Oya, the Yoruba orisha of storms and transformation. OYA AI is building climate world models trained on developing-world data to predict what generic global models miss: the precise track, intensity, and landfall impact of storms forming in data-sparse basins. The Caribbean loses 2.9% of GDP per major hurricane. Current warning lead time at landfall: 6 to 8 hours. OYA AI targets 72 hours of actionable lead time. We provide the future before it happens. Reinsurers, development banks, and emergency agencies pay for that.
Standard numerical weather prediction models run on global grids at 9–25km resolution. The Caribbean needs sub-kilometre nowcasting that accounts for island topography, shallow-water sea surface temperature anomalies, and orographic effects that generic global models smooth over. OYA AI's nowcasting engine ingests NOAA satellite data, ERA5 reanalysis feeds, and Maestro's proprietary Caribbean SST dataset to produce 72-hour tropical cyclone intensity and track forecasts at resolutions that matter for a 20-mile-wide island.
When a CAT-3 storm is 48 hours from landfall, every hour of preparation reduces recovery costs by an estimated $4 for every $1 spent. OYA AI's resilience automation platform manages the pre-event decision sequence that most Caribbean governments currently run through phone calls and spreadsheets: shelter capacity allocation, supply chain pre-positioning, utility shut-down sequencing, evacuation route optimisation, and coordination across overlapping agency jurisdictions. Automated, auditable, and always with a human decision-maker at the authorisation point.
Powered by Harmonics, OYA AI's disaster management agents are trained on Caribbean-specific emergency management protocols, CDEMA response frameworks, historical event data from 40 years of Atlantic hurricanes, and community-level vulnerability maps built from Maestro's Data Archaeology work. They function as always-available junior emergency management experts: briefing community leaders at 3am, running scenario analyses, tracking resource deployment, and maintaining institutional memory across the event lifecycle that human responders lose to exhaustion and staff turnover.
The reinsurance industry loses $90–130 billion annually to underpriced climate risk in emerging markets. The core problem is data: actuarial models for Caribbean and LATAM territories are built on sparse historical records and generic global climate models that do not capture the localised intensity patterns that determine whether a Category 2 becomes a $200M event or a $2B one.
OYA AI's nowcasting data and event records become a proprietary dataset for climate risk pricing that does not exist anywhere else. The social good mission and the commercial infrastructure are the same thing: better data produces better predictions, better predictions save lives and reduce losses, reduced losses are a commercial product that insurers, reinsurers, and development finance institutions will pay for.
Seven CDEMA member states are in active scoping conversations for OYA AI integration into national emergency management systems. The IDB and CDB have both identified AI-powered disaster preparedness as a priority investment area for the 2025–2030 cycle. Development finance is real revenue with social good baked in at the contract level.
| Revenue Stream | Buyer | Stage |
|---|---|---|
| Gov emergency management API | CDEMA members | Scoping |
| Climate risk data licensing | Reinsurers | Engaged |
| Development finance grants | IDB / CDB | Identified |
| NGO resilience platform | UNDP / OCHA | Conversations |
| Agent deployment: emergency mgmt | Nat. agencies | Early access |
Sea surface temperature is the primary fuel source for Atlantic tropical cyclones. OYA AI maintains a proprietary Caribbean SST anomaly dataset updated every 6 hours from MODIS and VIIRS satellite feeds, cross-referenced with historical intensification events from 1983–2024. This dataset does not exist in any public archive at the spatial resolution and update frequency OYA AI maintains.
Mountainous islands like Jamaica and Hispaniola generate orographic precipitation patterns that standard global models average away. OYA AI's topographic interaction models predict rainfall distribution and wind speed variance at sub-island scale, giving emergency managers the information they actually need: not how strong the storm is offshore, but what it will do when it crosses your coastline and hits your mountains.
Hurricane Maria intensified from Category 1 to Category 5 in 15 hours before hitting Dominica in 2017. Standard NWP models gave 6 hours of warning. OYA AI's rapid intensification signal model is trained specifically on Atlantic basin cases from 1990–2024, identifying the atmospheric and oceanic precursor patterns that precede explosive intensification with 72-hour lead times. This is the model that changes evacuation decision-making.
OYA AI is open to government partnerships,
development finance, and reinsurance data agreements.
15 years building AI systems before most companies knew they needed one. Founder of StarApple AI, the Caribbean's first AI company. A decade inside risk management at the highest levels, building the models that generated billions in revenue for banks and financial institutions. Now building the infrastructure for the half of the world that AI has never seen.
Adrian Dunkley has spent 15 years building AI systems from first principles. Not prototypes. Not demos. Production systems: credit risk models that ran on live bank portfolios, fraud detection architectures that processed millions of transactions, actuarial engines that priced insurance products across regional markets. The kind of work where a model error is not a sprint ticket. It is a financial consequence measured in millions.
A decade of that work was inside risk management. He built the quantitative models that generated billions in revenue for Caribbean and regional financial institutions, working at the intersection of applied mathematics, statistical inference, and domain expertise that no textbook prepared him for. Physics-informed modelling, Bayesian updating, causal inference applied to real credit portfolios with real default data. That grounding is what separates Maestro AI Labs from companies that discovered AI last year.
In 2018, he founded StarApple AI, the Caribbean's first AI company. Not because there was a clear market for it. Because the region needed it and no one else was going to build it. Seven years later, StarApple AI has trained hundreds of Caribbean organisations, advised CARICOM governments on AI policy, and produced the Caribbean AI Innovator of the Year who is now building the infrastructure those governments need next.
Maestro AI Labs is the convergence of that entire trajectory: the risk modelling rigour, the regional data knowledge, the institutional relationships, and the fifteen years of understanding that AI without context is just expensive noise.
Not research AI. Production AI. Credit scoring models deployed on live portfolios. Fraud detection systems running in real time. Natural language systems built before transformer architecture existed as a public concept. The practical difference between a model that works in a notebook and one that works in a bank is something you learn by breaking the latter in production. Adrian has broken enough to know exactly how to build it correctly.
A decade of applied risk management means understanding the specific failure modes that matter: model drift, data leakage, population shift, regulatory constraint. The quantitative models Adrian built for Caribbean and regional banks generated measurable revenue and managed measurable risk. That experience is the entire foundation of how Maestro AI Labs thinks about model risk, human oversight, and what it means to deploy AI in high-stakes financial environments.
The OYA AI nowcasting engine is not a black-box pattern matcher applied to weather data. It is a physics-informed system that encodes the actual atmospheric and oceanic dynamics of Caribbean storm formation. Adrian builds AI models the way a scientist builds them: from causal understanding first, statistical fitting second. That discipline produces models that generalise correctly when the distribution shifts, rather than failing silently when conditions differ from training data.
Investors, partners, and collaborators.
The conversation starts here.
If you are a financial institution seeking global credit intelligence, a government building context-aware policy tools, an enterprise deploying agents that need to work in Lagos as well as London, or a research institution that understands the stakes of training on half the world. Start here.