The Pipeline Starts at Five: Inside the US$1 Million Genius Project AI Bootcamp

Key takeaways

• The Genius Project 2026 enrolled over 200 participants aged 5 to 18 across the Caribbean and charged families nothing.
• US$1 million in cash and prizes was awarded across a single month, funded entirely by sponsors.
• The curriculum moves deliberately from AI tools into machine learning, statistics, mathematics, teamwork and problem definition.
• Teams built working models across crime and community safety, poverty and access, sport, and AI ethics.
• Parents ran a parallel track on safe AI use, critical thinking, and recognising AI slop and misinformation.
• Completion across all programme areas currently stands at roughly 15 percent, published openly.

We wrote earlier this year that the Caribbean cannot import AI sovereignty and cannot fly it in on a consulting contract that leaves when the invoice is paid. The IMPACT AI Lab is our answer to that at the university end: take strong graduates, put them on shipping products under senior review, and keep the talent and the intellectual property in-region.

That answer has a limit. IMPACT starts with people who already made it through a technical degree. The region loses far more people before that point than after it, and it loses them for reasons that have nothing to do with salary. They were never shown that this work was for them, or they hit the mathematics without anyone explaining what it was for.

The Genius Project works on that earlier stretch, and maestro sponsored the 2026 programme with technical mentorship and lab time for the machine learning tracks.

What one month produced

The curriculum goes past the tools

Most youth AI programmes stop at tool use. A student learns to prompt a chatbot, produces something that looks impressive, and leaves believing they understand AI. They understand an interface, which is a different thing and a far less durable one.

The Genius Project treats tools as the entry point and then moves the ground. Students went from prompting a model to understanding what a model is: training data, features, labels, and the difference between a system that learned a pattern and one that memorised an answer sheet. That requires statistics, and statistics requires mathematics. Distributions, probability, measurement error, and the reason a single result tells you almost nothing. For the older cohorts it meant Python, notebooks, and the ordinary experience of code that runs correctly and returns the wrong number.

Anyone who has read the PROMPTICA curriculum will recognise the bias. Evaluation is the spine. The gap between a demo that impresses and a system that survives real users is almost always a gap in how rigorously the team measures its own output, and the earliest place to install that habit is childhood.

Four problem areas, real models

Crime and community safety. Teams examined where incidents cluster, how gaps in reporting distort what the data appears to say, and what a model can responsibly claim about a place or a person. Several groups worked out for themselves that a predictive model built on incomplete crime data largely predicts where the reporting is. That is a finding professional teams reach late and expensively.

Poverty and access. Household budgeting tools, food price tracking, and matching people to services they qualify for but do not know exist. The wall these teams hit, that Caribbean household data is thin and scattered, is the same wall Credit Garden hits every day when scoring thin-file customers.

Sport. Football and track data proved the strongest on-ramp to machine learning available. Students already had domain intuition, which meant they could immediately tell when a model was producing nonsense. Most beginners cannot, and that inability is exactly how bad models reach production.

Ethics and responsible AI. Every team had to state who their system could fail, what data it should never hold, and what they would say to a person the model got wrong. A build requirement, not a lecture module.

Across all four areas the students built actual machine learning models. Systems that took input, produced output, and could be demonstrated to be wrong.

The parents trained too

A child who understands AI better than every adult in the household is not in a safe arrangement. Parents ran their own track alongside their children: account and privacy settings, what a chatbot retains, what should never be pasted into one, and how to recognise a website built to harvest information. Then judgement, which is the harder half. Telling a generated image from a photograph. Checking a claim before forwarding it. Recognising AI slop, the fluent and confident text that happens to be wrong, and understanding that the fluency is the trap rather than the reassurance.

The bar was modest and specific: a parent should be able to sit beside their child, look at the work, and ask one question that improves it.

The completion number, published

The programme publishes that figure rather than reporting enrolment alone, which is the sector norm and a habit that has made regional education data close to useless for comparison. Fifteen percent across a month-long technical curriculum spanning several countries is strong. It is also well below where it should sit.

The drop-off has two known causes. The first is the transition from tools to mathematics, which is where any technical curriculum loses people, and which is a design problem. The second is infrastructure: unreliable connections and nowhere quiet to work. That second one is not a curriculum problem, and it is the clearest illustration we have seen of how a connectivity gap converts directly into lost technical capacity. The 2027 design targets both with shorter mathematics modules, offline-capable materials, and a direct call to any participant who goes quiet for more than three days.

The hackathon

The month closed with a final hackathon. Teams presented to judges, defended their builds, and answered for their design choices. Congratulations to the winners, and to every team that presented at all. Defending a technical build in front of a panel of adults is difficult at thirty. A number of these presenters were not yet thirteen.

Why maestro funds the early end

Our position on Caribbean AI has been consistent. Models are a download away. The people who can wire one into a working product for a Kingston credit union or a Port of Spain clinic are not, and that shortage is what actually caps the region's AI ambitions.

IMPACT works on the last few metres of that pipeline. The Genius Project works on the first several kilometres, and the compounding is entirely in the early section. A nine-year-old who learns this year that a machine can be confidently wrong becomes a twenty-year-old who evaluates models properly, which is precisely the discipline PROMPTICA spends weeks trying to install in adults who never learned it.

The region cannot recruit its way out of a talent shortage against Toronto, London and New York salaries. It can grow its own, and the growing starts earlier than anyone budgets for.

Who funded the 2026 programme

The Genius Project charges families nothing, which works only because organisations across the region contributed cash and in-kind support. The 2026 sponsors were StarApple AI (prize funding, instructors and curriculum), maestro AI Labs (technical mentorship and lab time), the Caribbean AI Association (regional backing and CARICOM reach), 14West (hackathon and prize pool), AI Trinidad and Tobago (delivery across the twin islands), Orbital Brand Science (in-kind support and family outreach), and Adrian Dunkley personally.

To every sponsor, judge, volunteer instructor and parent who gave up a month of evenings: thank you. To the students: you did the hard part.

Families can register for the next cohort at beagenius.org. Tuition-free, from age five, virtual for participants outside Jamaica, no prior coding experience required. Organisations that want to sponsor a cohort can reach the programme through the same site.