Building for the Last Mile: Why AI in Education Won't Scale Until It Meets Learners Where They Are

A conversation with Gabriel Boluwatife Adegboye — product leader in EdTech and AI — on scratch cards, virtual reality nursing labs, and what happens to product management when the AI writes the first draft of the code.

By Louise Servoin · 2026-09-29 · 9 min read

## Gabriel Boluwatife Adegboye – Senior AI Product Manager at Miva Open University; Former Product Lead at Stakestack AI; Former Lead Product Manager at Miva Open University (uLesson Group)

Botany graduate turned product leader. Obafemi Awolowo University and Utiva Product School alumnus, currently pursuing an MSc in AI Engineering at Quantic School of Business and Technology. Has shipped EdTech, payments and AI learning products across West, East and Southern Africa, as well as for diaspora users in the United States and the United Kingdom. Global Youth Ambassador for Theirworld and alumnus of the School of Politics, Policy and Governance. Based in Abuja, Nigeria.

Most conversations about AI in education start with the model. Gabriel Adegboye starts somewhere else entirely: with K12 students or their parents seeking high quality educational content outside of the regular classroom, with a parent who won’t engage with any form of payment online due to lack of trust, a teacher on an unreliable mobile connection, and a post-secondary school student who couldn’t get a spot in the limited admission spots across Nigerian universities.

His path to that perspective is unusual. He spent four years at Obafemi Awolowo University studying botany, specializing in plant molecular genetics, running a morphological and molecular genetics study on three species of Boerhavia, before pivoting into software development, then into product management via the Utiva Product School — a transition he made while commuting across state lines from Oyo to Lagos to attend classes in person.

From there: four years at the uLesson Group under founder Sim Shagaya, first on the K-12 platform, then as Lead Product Manager starting up the digital foundations of Miva Open University in its first year of existence. Then teacher professional development at Instill Education in South Africa, and AI-native learning management at Stakestack AI and now back at Miva Open University as a Senior AI Product Manager.

We sat down with Gabriel to discuss why the hardest problems in EdTech are rarely pedagogical, what virtual reality actually solved for his nursing students, and why the hardest design decisions in an AI learning product have nothing to do with the model.

## From the Lab Bench to the Roadmap

You started in Botany — plant anatomy, physiology, ecology, genetics, biometrics. That is not the usual on-ramp to product management. What carried over?

More than people expect. In the lab, you learn to isolate one variable at a time and resist the temptation to explain a result before you have the data to support it. My final-year work was a molecular and morphological study of three Boerhavia species — the kind of work where you track variables across control and treatment groups at different dosages, and the discipline is entirely about not confusing a pattern with a cause.

That is the same muscle you use in product. An A/B test is a controlled experiment. A retention curve is a dose-response relationship in disguise. What science gave me was a healthy suspicion of my own conclusions, challenging assumptions and testing hypotheses, which is useful in a job where everybody has an opinion about what users want.

And the pivot itself?

The pivot came from a mix of failure, passion and interest, honestly. I worked as a web developer for about two years, and in parallel I tried to build and launch my own products. They did not get traction. What I eventually understood is that I was very good at executing specifications and very poor at questioning whether the specification was worth building. Technical execution without a validated product vision is expensive motion.

That is what took me to Utiva in 2019. I was commuting between states for physical classes — design thinking, product design, agile execution, product leadership. Funto Akinbisehin's session on product leadership was the one that reframed everything for me. After that, I did not wait for a recruiter. I cold-emailed founders and product leaders, until Desiree Craig gave me a chance, and that was how I joined the uLesson Group on the 26th of October 2020.

## Designing for Infrastructure, Not for Slides

At uLesson you were shipping across West and East Africa, plus the US and UK. What broke first?

Payments. If you ship a globally standard checkout — credit/debit cards, one currency, one flow — you will lose a very large share of your users at the payment screen. Not because parents don't want to pay, but because the rails just can’t serve them (in a one-size fits all approach), or they can’t find their most preferred payment methods. Mobile money is the pre-dominant method in East Africa for example, and in certain regions of some markets, cash is still the default because trust in formal banking is limited.

So we stopped trying to make users fit the checkout and made the checkout fit the users. We shipped localized collections so parents could pay in their own currency through the mobile money or bank transfer method they already used. For the cash-based segment, we designed a scratch card: you buy a physical card with cash from a local agent, you type the code into the app, you're subscribed. No bank account, no card, no drop-off.

And the content side of expansion?

Same principle, different constraint. When we moved beyond Nigeria and Ghana into East Africa and the diaspora markets, we hit the fact that every country has its own curriculum, its own grade naming, its own exams — WASSCE, GCE, UTME. Building a separate platform per country was financially and operationally impossible.

So we built content localization as a layer on top of the existing CMS, working with curriculum subject-matter experts. One backend, dynamically presenting the right content for the right market. That is the decision that made pan-African expansion viable rather than theoretical.

Was there a hardware dimension too?

Yes — the education tablet, also pre-loaded with lessons for offline learning, because bandwidth is a real constraint for a large share of students. My job there was the commerce layer: a checkout flow optimized to reduce drop-off, and a system that captured incomplete checkouts as leads for the customer service team instead of letting them evaporate. And underneath all of it, server-to-server event tracking across our analytics stack, because you cannot improve a funnel you cannot see.

"The technology is almost never the hard part. The hard part is the last mile — the parent who doesn’t trust the online banking or payment system, the student/teachers on a low-bandwidth connection, the student who couldn’t secure a tertiary institution admission seat, or the students who needed to combine work with studies." — Gabriel Boluwatife Adegboye

## Virtual Reality and the Clinical Placement Gap

At Miva Open University you set up a VR simulation lab for nursing and public health. Why VR, and why then?

Because the bottleneck in nursing education is not lectures. It is clinical hours. Programs everywhere struggle to secure enough real hospital time for students — patient privacy, limited space, and the simple fact that you cannot hand a complex emergency to a novice without risk. In Nigeria that scarcity is compounded by infrastructure gaps.

VR does not replace the ward, but it changes what a student can practise before they get there. In simulation, a student can misjudge a dosage or miss a diagnosis and nobody is harmed. That psychological safety is the pedagogical point: you learn most from the error you were allowed to make. Beyond procedures, the scenarios force real communication — structured handoffs, talking to a virtual patient — which a plastic mannequin cannot teach you. And you can expose students to rare, high-stakes cases they might never encounter in a standard rotation.

What made it hard as a product problem?

As a university in its founding year, building simulation hardware and software in-house was completely out of the question—we simply did not have the internal engineering bandwidth to commit to a multi-year deep-tech build. That meant sourcing a commercial vendor, which brought a different set of headaches.

The challenge was finding a provider whose content aligned strictly with local clinical accreditation guidelines without forcing us into prohibitive license fees. We had to evaluate providers rigorously on pedagogical quality, ease of LMS integration, offline/hybrid usability, and cost per seat.

Ultimately, managing those trade-offs meant prioritizing modularity and clinical relevance over flashy graphical fidelity. We chose a partner whose platform offered us relevant varieties of scenarios, ensuring our nursing students gained practical, accredited skills within our operational constraints.

## When AI Becomes the Product

At Instill Education and then Stakestack AI, the work shifted toward teachers and AI-native learning. What changed in your thinking?

Instill was a reminder that the teacher is the highest-leverage user in any education system. We were building a mobile-first professional development platform used by educators across South Africa, Kenya, Ghana and Nigeria. Mobile-first there is not an aesthetic preference — it is a constraint. Older devices, weaker connections, teachers using the product between classes. You design for that or you don't get used.

Stakestack is the other end of the spectrum: an AI-powered learning management system for professionals, corporations and institutions. The premise is that the one-size-fits-all training module is finished. Personalized learning paths, text-to-video generation, automated curriculum design — the content layer becomes dynamic rather than fixed.

What is the piece of that you find most interesting?

Assessment. I built and launched an AI-powered cognitive test — an assessment that tries to determine how well a learner actually understands a subject, rather than whether they can recognize the right option out of four. Multiple choice measures recall under constrained conditions. It has survived because it is easier and maybe cheaper to grade, not because it is a good instrument. If AI can make richer assessment cheap, that is a bigger change to education than any content generation feature.

## The Vibe Coding Question

You describe yourself as proficient in "vibe coding." For a product manager, what does that actually change day to day?

It collapses the distance between having an idea and seeing whether it holds up. The term came from Andrej Karpathy in early 2025 — using natural language to get a large language model to generate, refactor and debug code. By now it is a normal part of how software gets built, and the tooling has split into AI-assisted IDEs for engineers, generators for non-technical founders, and narrow specialist tools.

For me, it means I no longer have to spend engineering capacity to find out whether a concept is worth building. I can put a working prototype in front of a stakeholder or a user, watch what happens, and only then write the PRD that commits a team to a quarter of work.

And the risks?

They are real, and I'd rather name them than pretend otherwise. Generated code can carry technical debt and security exposure if nobody reviews it. The honest framing is trust but verify: let the model write the boilerplate, keep human ownership of business logic, security and architecture. It is an accelerator for prototyping and validation. It is not a licence to push unreviewed/untested code into production.

That is also part of why I went back to school. I'm doing an MSc in AI Engineering at Quantic — Managing AI engineering, AI-assisted software development, AI engineering techniques and architecture, Model fine-tuning, Microservices architecture, agents and multi-agent systems, cloud architecture and scaling, AI and organizational transformation, plus the governance and innovation-leadership side. I'll also be getting into a little of the math behind AI. If you are going to lead AI products, you need to understand the failure modes of the thing you're selling.

## The Question of the Future

As AI commoditizes "average" knowledge, what is the human skill in your field that you don't think an algorithm replaces?

Knowing what is actually true on the ground. A model can tell you the optimal checkout flow based on global data. For example, It will not tell you that in a particular market, the parent pays in cash at a kiosk because that is the rail their local economy actually runs on, and that the entire product depends on respecting that.

That kind of knowledge comes from proximity — from talking to users in their context, not from a dataset about them. Operational empathy is the term I'd use. It is why solutions built somewhere and shipped elsewhere can fail: they are technically excellent and contextually wrong.

I'd add taste, or judgment. Deciding what deserves to exist is still a human act. My own mission statement is "tech for humanity" — technology as a lever for reducing everyday friction and advancing the continent, not as an end in itself. No model gives you that; it only executes against it.

## The Golden Thread

We often see a gap between what technology can do and what professionals actually adopt. How do we ensure AI becomes a 'co-pilot' that elevates human judgment rather than a tool that encourages intellectual passivity?

I think the answer is to design/build AI to make people think better, not think less.

The danger is not that AI will make people less capable. The danger is that we will build products that reward people for accepting the first answer the model gives them. If AI becomes an answer machine, intellectual passivity is almost inevitable. But if it becomes a thinking partner, one that asks you to explain your reasoning, challenges your assumptions, surfaces alternatives, and makes you confront uncertainty, it can actually raise the quality of human judgment.

That means we have to be deliberate about where we put the human in the loop. In education, for example, I don't want an AI tutor that simply gives a student the answer. I want one that can recognise when the student is struggling, ask the right questions, offer a useful hint, and then make the student do the cognitive work. The same principle applies to professionals. An AI product should not just say, "Here is the recommendation." It should help you understand why that recommendation was made, what assumptions it depends on, what could make it wrong, and what information you should consider before acting on it. This also shines light on the blind spots created by just taking whatever AI tells us.

I also think professionals need to maintain a healthy level of skepticism toward AI. The fact that a model can produce an articulate answer in three seconds does not make the answer correct. The human's role increasingly becomes knowing what questions to ask, how to ask the question (so that the model doesn’t give you an answer it thinks would make you happy) what evidence matters, what doesn't make sense in context, and when to challenge the model.

Ultimately, I don't think the future belongs to people who can do everything without AI. It belongs to people who know when to use AI, how to challenge it, and when their own judgment should overrule it.

The best AI should leave you more capable after using it, not totally dependent on it.

Tags: AI in Education, EdTech, Product Management, Africa, Personalized Learning, Assessment, VR in Healthcare Education, Vibe Coding