Credit Garden: How Physics-Informed AI Is Rebuilding Caribbean Credit

Key takeaways

• About 60% of Caribbean adults are thin-file: legacy bureaus return blank rather than a score, and a blank is treated as a decline.
• Credit Garden scores borrowers on consented mobile money, utility, airtime, and small-business cash flow instead of mortgages and credit cards they never held.
• Its physics-informed “equity of borrowers” model maps where a borrower’s finances are heading, not just where they stand today, so seasonal incomes are priced to their real path.
• Indicative pilots show defaults among previously unbanked borrowers falling by more than half, roughly twice as many approvals, and lower recovery cost.
• Because the model reads behaviour over months rather than a single self-reported snapshot, it is harder to game than a static score.
• Lenders can run Credit Garden in shadow mode first, comparing approvals, pricing, and realised defaults against their current book before changing any policy.

A woman in Montego Bay runs a hair salon, pays her light bill on time for nine years straight, tops up her phone every week, and moves money through a mobile wallet to three suppliers. She has never held a credit card or a mortgage. To a legacy credit bureau, she does not exist as a borrower. Her file is too thin to score, so the model returns nothing, and the lender treats nothing as a no. Multiply her by tens of millions and you have the Caribbean’s credit problem in one sentence: the people the system cannot see are not high risk, they are unmeasured.

This is the gap Credit Garden was built to close. It is maestro’s AI credit product, developed with StarApple AI, and its premise is blunt. The borrowers legacy bureaus reject are mostly creditworthy. The scorecards are wrong, not the people.

The thin-file problem is a measurement problem

Roughly 60% of Caribbean adults carry a thin or non-existent credit file. That figure is an artefact of the measuring instrument rather than the population it scores. Classic credit scores were designed in economies where most adults hold mortgages, revolving credit cards, and auto loans, and where decades of repayment data sit in a central bureau. Feed that scorecard a borrower with none of those products and it does not return “risky”, it returns blank.

Blank is the trap. A bank cannot lend against a blank, so it declines, and the decline keeps the borrower from building the very history the bureau wanted to see. The result is a closed loop that locks out the informal economy, which in much of the Caribbean is the real economy: market vendors, taxi operators, hairdressers, fishers, and one-person construction outfits that run on cash and mobile money. These are not edge cases. They are the majority.

What Credit Garden actually reads

Credit Garden starts from the data Caribbean people generate every day, the trail legacy bureaus throw away. Mobile money flows show income rhythm, supplier relationships, and how a household absorbs a shock. Utility records, the light bill and the water bill, are one of the strongest honest signals of stability anyone has: a person who keeps the power on for a decade is telling you something a FICO-style score never asked. Airtime top-up patterns track liquidity at the weekly level. For small businesses, transaction cash flow shows whether revenue is growing, seasonal, or stalling.

None of this is exotic data. It is ordinary, and that is the point. The signal was always there. What changed is that generative AI can now read messy, semi-structured, multi-source financial behaviour the way an experienced loan officer reads a community, and do it at the scale of a national portfolio. Credit Garden uses that capability to turn a thin file into a thick picture, drawn from sources the borrower already controls and consents to share. It sits in the same family as the other Caribbean-built systems in maestro’s product line, and it is part of the wider case that the region can ship its own AI rather than rent it, a theme we trace in five AI innovations putting the Caribbean on the global tech map.

Physics-informed scoring: a trajectory, not a snapshot

Here is where Credit Garden parts company with every alternative-data scorecard that came before it. A standard score, classic or alternative, is a snapshot. It compresses a borrower into a single number at a single moment and asks: how risky is this person today? That framing throws away the most important thing about a human being’s finances, which is that they move.

Credit Garden treats a borrower’s financial life as a system with momentum. We call the underlying method an “equity of borrowers” model, and it borrows its mathematics from physics. In a physical system, you do not predict where a body will be by looking only at its current position. You account for its velocity, the forces acting on it, and the path it has been following. Credit Garden applies the same logic to money. It models income velocity, the direction household equity is heading, the shocks a borrower has already absorbed and recovered from, and the forces, seasonal, sectoral, regional, pushing the trajectory one way or another.

The output is a forward map of where a borrower’s finances are likely to travel over the life of a loan, generated before the loan is written. A snapshot score might decline a fisher in the off-season because this month looks weak. A trajectory model sees the annual cycle, the recovery pattern, and the upward slope, and prices the loan to the real path rather than the worst week. That difference is why physics-informed scoring approves people a static score rejects, while keeping losses down.

The results, and why lenders care

The case for Credit Garden is not charity, it is arithmetic. maestro’s own homepage carries a lender testimonial that captures it: among unbanked borrowers, defaults dropped by more than half, more loans were approved, and recovery costs fell. Those three numbers move in the same direction for a reason. Better measurement does not force a trade-off between approving more people and losing less money. It improves both at once, because the losses in a thin-file portfolio come mostly from guessing, and Credit Garden replaces the guess with a model.

Read the chart at the top of this piece again. A legacy bureau, forced to score a thin-file pool it was never built for, runs default rates near 18%. Credit Garden, scoring the same borrowers on behaviour and trajectory, brings that closer to 8%. For a lender, that is the gap between a segment that loses money and a segment worth competing for. Lower defaults, lower recovery spend, more good loans on the book. The unbanked stop being a risk to avoid and become a market to win.

Why a trajectory model is harder to game than a score

Any scoring system that touches money gets attacked. The moment a number controls who gets a loan, people start engineering the number, and alternative-data scores are easy targets because so many of them rest on a thin set of self-reported or one-off inputs. Stuff a wallet with a few large transfers the week before applying, borrow a friend’s phone with a clean top-up history, and a naive behavioural score lights up green. Static scores reward whoever can dress up a single moment.

Credit Garden is built to resist exactly that. Because the “equity of borrowers” model reads velocity and trajectory across months, a sudden cosmetic spike does not look like strength, it looks like an anomaly against an otherwise flat path, and the physics of the model treats it as noise rather than signal. Genuine creditworthiness in this framing is a consistent direction of travel: income that recovers after a shock, supplier relationships that persist, a household that keeps the lights on through a rough quarter. Those patterns are expensive to fake because faking them means actually behaving like a reliable borrower for a long time, which is the behaviour the lender wanted in the first place. Gaming a snapshot is cheap. Gaming a trajectory means becoming the thing you were pretending to be.

That robustness against manipulation matters most precisely where inclusion matters most, in markets with limited formal verification infrastructure. A model that can be tricked by a clever applicant pushes lenders back toward the conservative declines that created the thin-file problem. One that reads sustained behaviour lets a lender say yes to real borrowers without opening the door to fraud. This is the same standards-first instinct behind Section 9, maestro’s AI safety lab: a system trusted with consequential decisions has to hold up when people try to bend it.

Remittances and diaspora income as a credit signal

For a large share of Caribbean households, the most reliable line of income is not a local paycheck, it is the transfer that lands every fortnight from a relative in Brooklyn, Brixton, or Toronto. Remittances are among the steadiest inflows in the region’s economy, and in several Caribbean nations they run well into double digits as a share of GDP. To a legacy bureau, that income is invisible: it is not a salary, not a credit line, and not anything the scorecard knows how to read. A household supported by a decade of dependable diaspora transfers can still come back unscoreable.

Credit Garden treats consented remittance flows as what they are: a durable, trackable income stream with its own rhythm and resilience. The model can see how regular the transfers are, how the household deploys them, and how the pattern holds through a sender’s job change or a currency swing. Diaspora income often behaves more predictably than informal local earnings, because the sender’s commitment is personal and long-running. Reading it properly lets a lender extend a first mortgage or a business loan to a family the formal system had written off, on the strength of money that was always there. It also keeps more of that capital working inside the region rather than sitting idle in a wallet, which is the kind of compounding effect we explored in our look at AI and the Caribbean economy.

What it unlocks: wealth, not just access

Credit is the on-ramp to everything else. A salon owner who can borrow against her real cash flow buys a second chair and hires an apprentice. A farmer who can finance inputs at a fair rate plants more and sells more. A family that can access a first mortgage starts building the asset that, held across a generation, becomes inheritance. Shut people out of credit and you do not only deny them a loan. You deny them the compounding that turns income into wealth.

That is the deeper stake in fixing the scorecard. The Caribbean does not have a shortage of hard-working, creditworthy people. It has had a shortage of instruments that can see them. When Credit Garden brings the thin-file majority into fair credit, the effect runs past any single loan: more small businesses survive their second year, more households move from renting to owning, and more capital circulates inside Caribbean economies instead of leaking out in declined applications and informal-lender interest. Financial inclusion done well is generational wealth with a slow fuse.

What it means for SMEs and women-owned businesses

The thin-file penalty falls hardest on small and medium enterprises, and harder still on the women who run a large share of Caribbean micro-businesses. A salon, a catering operation, a small farm, a roadside retailer: these firms live on cash and mobile money, rarely carry audited statements, and almost never hold the kind of collateral a traditional underwriter asks for. The owner is often personally creditworthy and the business demonstrably viable, yet both come back unscoreable, so growth capital flows to the firms that already had access. The gap is not a rounding error in Caribbean development, it is one of the main brakes on it.

Credit Garden reads an SME the way the business actually operates. Transaction cash flow shows whether revenue is growing, seasonal, or stalling. Supplier payment patterns reveal reliability that no balance sheet captures. The trajectory model can tell the difference between a catering business that dips every January and recovers by carnival season and one that is genuinely sliding, and price the loan to the real cycle rather than the slow month. For women-owned firms in particular, that shift matters: it replaces the subjective judgement and collateral requirements that have historically tilted against them with a measurement of how the business performs. When the instrument reads behaviour instead of assets, a market vendor with eight years of steady mobile money flow finally looks like what she is, which is a good loan. That widening of who can build a business is part of the broader mission behind the IMPACT AI Lab, and a practical answer to the question of who actually benefits when the Caribbean builds its own AI.

How lenders and governments can pilot it: the first 90 days

Credit Garden is built to run inside an existing lender, not to replace one. A bank, credit union, microfinance institution, or development agency starts with a defined segment, usually the thin-file applicants it currently declines, and runs Credit Garden in shadow mode alongside its current process. The model scores the same applicants on consented behavioural data, and the lender compares approvals, pricing, and realised defaults against its existing book before changing a single policy. The evidence comes first.

The first 90 days follow a clear arc. Weeks one to three are integration and consent: connecting the data sources the lender already touches, mobile money, utility, and transaction feeds, and standing up the consent flow so borrowers opt in to share their own behaviour. Weeks four to eight are the shadow run, where Credit Garden scores live applicants in parallel without making a single lending decision, building a side-by-side record against the lender’s current scorecard. The final stretch is calibration and review: comparing the two on approval rate, pricing, and early repayment behaviour, then setting the policy thresholds the institution is comfortable with. Nothing changes on the production book until the lender has its own numbers in hand. A pilot does not ask a credit committee to trust a vendor’s slide deck, it asks them to read their own portfolio.

Data privacy and regulation sit at the centre of that design, not at the edge. Credit Garden runs on consent: a borrower chooses to share the financial trail they already generate, in exchange for a fair shot at credit the old system denied them. That consent-first model fits the data-protection regimes maturing across CARICOM, including Jamaica’s Data Protection Act and the frameworks regulators in Trinidad and Tobago, Barbados, and across the region are building out. Because the model can explain which behavioural factors drove a decision rather than hiding behind an opaque score, it gives lenders and supervisors an audit trail, which matters as Caribbean central banks weigh how alternative data should be governed. We treat that as a feature: the same explainability and safety discipline runs through maestro’s wider work, including PROMPTICA and the standards we build in our research practice. Used well, alternative data is not a privacy risk to manage, it is borrower-controlled information that finally counts in the borrower’s favour.

For governments and regulators, the same engine supports financial-inclusion mandates with measurement rather than slogans. A central bank or development ministry can pilot Credit Garden across a portfolio of national lenders, watch inclusion and default rates at the same time, and set policy against real outcomes. The regional model already spans 14 Caribbean nations, so a pilot in one market draws on context from the whole region. If you run a lending book or a financial-inclusion programme and 60% of your potential market is currently invisible, Credit Garden is the instrument that lets you see it.

Frequently Asked Questions

What is Credit Garden?

Credit Garden is maestro AI Labs’ AI credit product, built with StarApple AI. It scores thin-file Caribbean borrowers, the roughly 60% of adults legacy bureaus cannot rate, using everyday financial behaviour and a physics-informed model that maps a borrower’s financial trajectory before a loan is approved. The goal is more approvals with lower losses.

What does “thin-file” mean, and why are so many Caribbean adults thin-file?

A thin-file borrower has too little formal credit history for a traditional bureau to score. Classic scorecards were built around mortgages, credit cards, and auto loans held over many years. Much of the Caribbean economy runs on cash and mobile money in the informal sector, so a large share of adults, around 60%, never generate the records those scorecards expect, even when they are reliable payers.

What data does Credit Garden use?

Credit Garden reads consented, everyday financial signals: mobile money transaction flows, utility bill payment records, airtime top-up patterns, and small-business cash flow. Generative AI interprets this messy, multi-source behaviour at portfolio scale, turning a thin formal file into a rich, accurate picture of how a borrower actually manages money.

How is physics-informed scoring different from a normal credit score?

A normal score is a snapshot: one number at one moment. Credit Garden’s “equity of borrowers” model treats finances as a system with momentum, accounting for income velocity, recovery from past shocks, and the forces shaping a borrower’s path. It projects where the borrower’s finances are heading over the life of the loan, so seasonal or cyclical incomes are priced to their real trajectory rather than their weakest week.

What results does Credit Garden deliver for lenders?

Indicative pilot outcomes and maestro’s lender testimonial point the same way: defaults among previously unbanked borrowers fell by more than half, roughly twice as many applicants were approved, and recovery costs dropped. Better measurement raises approvals and cuts losses at the same time, because most losses in a thin-file book come from guessing rather than from the borrowers themselves.

How can a lender or government pilot Credit Garden?

Start in shadow mode. Credit Garden runs alongside a lender’s current process on a defined thin-file segment, scoring the same applicants on consented behavioural data so the institution can compare approvals, pricing, and realised defaults before changing policy. Governments and regulators can pilot it across national lenders to track inclusion and risk together. Visit the products page or get started to scope a pilot.

What do the first 90 days of a Credit Garden pilot look like?

Roughly three to four weeks for data integration and standing up the borrower consent flow, then four to five weeks of shadow scoring where the model rates live applicants in parallel without making decisions, then calibration and review against the lender’s own approvals, pricing, and early repayment data. No production lending decision changes until the institution has its own side-by-side numbers in hand.

Is alternative-data credit scoring allowed under Caribbean data-privacy law?

Credit Garden is built on borrower consent, where a person opts in to share the financial behaviour they already generate. That design fits the data-protection regimes maturing across CARICOM, including Jamaica’s Data Protection Act. The model can explain which behavioural factors drove a decision, giving lenders and regulators an audit trail rather than an opaque black-box score.

Can Credit Garden be tricked or gamed by applicants?

It is far harder to game than a static score. Because the model reads income velocity and trajectory across months rather than a single moment, a cosmetic spike such as a one-off large transfer reads as an anomaly rather than strength. Faking a genuine upward trajectory means actually behaving like a reliable borrower over a sustained period, which is exactly the behaviour a lender wants.

How does Credit Garden help women-owned and small businesses?

Small firms and the women who run much of the Caribbean’s micro-business sector rarely hold audited statements or collateral, so traditional underwriting declines them even when they are viable. Credit Garden reads transaction cash flow and supplier payment patterns instead, pricing a loan to a business’s real seasonal cycle. That replaces subjective, collateral-heavy judgement with a measurement of how the business actually performs.