José Guadalupe Posada, Cover for 'La Rumba: Coleccion de Canciones Modernas para el Presente Año 1903', a conductor and a drummer performing on stage for an audience, ca. 1903. The Metropolitan Museum of Art, public domain.
In April, the Jamaica Observer ran a story headlined "Jamaican AI loading." The key fact sat well down the page. A team of three people, working with my brother Nicholas and me at Maestro AI Labs, was building Jamaica's own AI model for a fraction of what people assume that costs.
Building a large language model, the kind of AI behind ChatGPT, from the ground up, the way the big US labs do, is estimated to cost US$5 million to US$15 million before anyone uses it. For a Caribbean economy, that has kept sovereign AI in the "someday, when a government or development bank pays" category. Project Maestro shows a cheaper route.
How the team built it: adapt, do not rebuild
The team started from existing open models, which their makers publish for anyone to download and adapt, instead of training from zero. They removed what did not belong and retrained the result on approved Jamaican data: language, local context, regulation, and how business and government actually work here. That choice turned a year-long, multi-million-dollar build into three months at a fraction of the cost.
My PhD research in climate physics follows the same logic. The large climate models behind global policy are computationally enormous and built by wealthy nations. My work develops generative-AI climate models that aim to approach their accuracy at a fraction of the cost, so a small island state does not have to wait on another country's supercomputer to see a flash drought coming. Project Maestro applies that approach to language models: adapt what exists instead of trying to outspend Silicon Valley.
The core team was three people, working with a few AI agents, volunteers and young Jamaican developers. Small teams are what building national infrastructure in a small economy looks like, and the constraint pushed the team toward the cheaper method. Nvidia is providing GPUs, training infrastructure and help with capital-raising and go-to-market to the three-person team in Kingston.
What a sovereign model does that a general one does not
People often ask why the Caribbean needs its own models when ChatGPT and Claude work here. They mostly do, until a task depends on context they were never trained on:
- Patois and the way Jamaicans actually phrase questions.
- Informal credit and remittance relationships that never appear on a bank statement.
- Drought and hurricane patterns specific to small islands.
- Local regulation that a model trained on US and European law has no reason to know.
Control is the other reason. When a country routes all its important AI use through foreign vendors, it depends on decisions it did not make and cannot audit. I chair the Caribbean AI Risk Management Council, and that is the argument I make there. With its exposure to climate shocks, financial exclusion and thin institutional capacity, the region needs AI that works correctly on its own cases.
Credit Garden, from the same lab, applies the same idea to lending. It turns five years of corridor-level remittance data into a structured credit input for 18 Caribbean and Latin American economies. Remittances into the region run to billions of dollars a year, and for a household receiving money every month they are a strong sign of stability that Western credit models have ignored. In the company's validation testing across 12,000 historical loan outcomes, including remittance context adjusted scores by an average of 302 points with no increase in the default rate.
Project Maestro, By the Numbers
- $5–15MTypical cost to build a large language model from scratch
- 3Core team members building Project Maestro
- 3 monthsTime to do what normally takes about a year
- 18Caribbean and LATAM economies covered by Credit Garden's remittance data
- 302 ptsAverage credit score adjustment Credit Garden found with no rise in default rate
Why red-team testing comes before launch
Maestro AI Labs has a dedicated team trying to push Project Maestro into harmful or unethical output before anyone outside the company uses it. A model that fails during internal testing costs the team a week. One that fails in front of the agencies and businesses it was built for sets back the region's credibility on AI for years. That is also why every Caribbean government already has free access to TurtleBird, the AI safety toolkit I built through Maestro AI Labs.
Testing takes as long as it takes. The founders plan regional expansion and, eventually, a public listing to fund it. Neither is needed to make the main point: a sovereign model no longer requires a national champion or a government budget line in the hundreds of millions.
The open question I cannot answer yet is upkeep. Adapting an open model is cheap once; keeping it current as base models improve, laws change and new Jamaican data arrives is a running cost, and I do not yet know what that will be each year. Any government planning to rely on a sovereign model should ask for that figure before signing.
What this means for the rest of the region
At Caribbean AI Association events and in meetings with finance ministers, I am often asked whether a small country can build its own AI. Project Maestro is a working example. If three people and a retrained open model can reach red-team testing in three months, Trinidad and Tobago, Barbados, Guyana and the OECS states do not need sovereign wealth funds to start.
When I founded StarApple AI as the Caribbean's first AI company, there was no other example to point to. The region now has labs, funds, a safety council and a sovereign model built on a budget a mid-sized Caribbean business could raise. The point of building StarApple AI, Maestro AI Labs and the other ventures I have backed was that the next founder would not have to start from zero.
What to do next
- ICT ministries in other CARICOM states: list the three government tasks where general AI models fail on local context (for example, answering questions on national regulations), and use them as the test set for any sovereign or adapted model you consider.
- Data owners in government and business: audit which of your datasets you can lawfully use for training, and record the consent or legal basis for each, before any model work starts.
- Teams building a local model: start by adapting an open model whose licence allows your intended commercial use, and budget for a red-team phase before any external user sees it.
- Lenders in Jamaica and the region: run a back-test of your own past loans with remittance history added, and compare default rates before changing any credit policy.
- Every Caribbean government: take up the free TurtleBird access and run it against any AI tool your agencies already use.
Frequently Asked Questions
What is the difference between training a model from scratch, fine-tuning and retrieval?
Training from scratch builds a model's knowledge from billions of words and costs millions of dollars. Fine-tuning, which is what Project Maestro did in adapted form, retrains an existing open model on a smaller local dataset for a fraction of that. Retrieval leaves the model unchanged and feeds it relevant documents at the moment of each question, which suits rules and prices that change often. Most practical systems combine fine-tuning with retrieval.
What does red-team testing involve?
A team deliberately tries to make the model misbehave: produce harmful instructions, leak data, give biased answers or ignore its rules, using tricks such as role-play, disguised requests and long conversations. Each failure is logged, fixed and retested. The test set should include local cases, such as Patois prompts and Jamaican legal questions, because failures on those will not show up in generic benchmarks.
Can Jamaican personal data be used to train an AI model?
Only with a lawful basis under Jamaica's Data Protection Act, which came into force in December 2023. In practice that means consent or another basis the Act allows, clear purpose limits and security for the data. The Office of the Information Commissioner oversees compliance, and a project using personal data should document its assessment before training starts.
Are all open models free to adapt for commercial use?
No. Open model licences vary: some are fully permissive, others restrict certain uses, require attribution or set conditions for deployments above a stated number of users. Read the licence of the exact model version you start from, and get legal advice if a government or bank will rely on the result.
How would we know a sovereign model is better than ChatGPT for Jamaican work?
Test both on the same set of local tasks and score the answers blind. Include Patois questions, Jamaican regulations, local place names and realistic customer queries, and have Jamaican reviewers grade them without knowing which model answered. Publish the results so users can judge the claim for themselves.