Welcome to Issue #75 of TAIS, where every Friday we spotlight visionary changemakers reshaping Africa’s tech, data, and AI landscape, one breakthrough at a time.
Today we spotlight Aaron Charles, founder of MasomoGPT & UtuAfrica, whose path into African education technology began not in a boardroom but in a classroom in Karen, Nairobi, teaching AI to ten-year-olds every morning. His world operates where curriculum meets code, where a student typing a fraction question in Swahili and English in the same sentence (the way most Kenyan children actually think) exposes exactly what every foreign AI tool gets wrong: it treats her code-switching as an error to correct rather than a comprehension mode to understand. Not in the space where “AI for education” gets discussed as a philosophical question of access or sovereignty, but in the harder, less visible arithmetic underneath it: what it actually costs to run a model at scale on a KES 4,000 phone with intermittent 2G, and whether that cost makes the whole enterprise viable or impossible before a single lesson is taught.
Aaron is not digitising education. He is arguing that digitising it was never the point. His conviction, which runs through every answer in this conversation, is that Africa doesn’t need a translated version of Western EdTech, it needs intelligence built from scratch around how its students actually think, speak, and live economically. He refuses to let “personalisation” mean adjusting pace while keeping American examples and American assumptions intact, and he refuses to rent that intelligence from someone else’s API, training his own model because, as he puts it, the real question isn’t philosophical, it’s arithmetic.
His work with MasomoGPT is his answer to a question the industry rarely pauses to ask: not whether AI can reach African students, but whether the unit economics of reaching them will ever add up if the intelligence isn’t owned locally.
Q: You founded MasomoGPT with the ambition of using AI to transform education, but your work has since grown into a broader ecosystem that includes AI tutoring, digital books, educational media, and other learning tools. What first convinced you that Africa needed a fundamentally different approach to learning, rather than simply digitising existing education?
A: I teach children AI every morning in Karen, Nairobi. Ten-year-olds. And the thing that struck me early was not that these kids lacked access to information — they have phones, they have YouTube, they can Google anything. What they lacked was an intelligent system that understood their context.
I watched a student type a question about fractions using Swahili and English in the same sentence — what we call code-switching, which is how most Kenyan students actually think and communicate. Every foreign AI tool treated that as an error. It either ignored the Swahili or returned garbage. The student got a "correct" answer that was completely useless to her.
That was the moment. Digitising existing education means taking a textbook and putting it on a screen. That solves access, maybe, but it does not solve comprehension. A Kenyan student in Grade 6 does not think in abstract English. She thinks in the local context of matatu fares and maize prices and shillings. When she asks "Explain photosynthesis," she expects an answer that maps to her CBC curriculum — not a US fifth-grade science explanation with examples of oak trees she has never seen. Africa does not need digitised education. It needs intelligent education that understands the student — linguistically, culturally, economically, and pedagogically. That requires building from scratch, not importing and translating
Q: Personalised learning is often presented as one of AI's greatest promises in education, yet it can mean very different things in practice. What does truly personalised learning look like for students in African education systems, and where do you think current approaches still fall short?
A: Truly personalised learning in an African context means three things most platforms miss entirely.
First, linguistic personalisation. A student in Kisumu thinks differently from a student in Mombasa — not because one is smarter, but because their cognitive frameworks are shaped by different mother tongues, different code-switching patterns, different ways of structuring an argument. Our model handles seven African languages in production today — English, Swahili, Sheng, Dholuo, Kalenjin, Kamba, and Luhya — not as translation layers but as native comprehension modes. When a student mixes languages mid-sentence, the AI responds in kind. That is personalisation. A Swahili toggle on a US model is not.
Second, economic personalisation. Our students are on KES 4,000 phones with intermittent 2G connectivity. Personalisation means the platform must work offline, must render on low-end devices, must cost less than a daily mandazi to access. If your "personalised" AI requires a stable 4G connection and a modern smartphone, you have personalised for the top 10% and excluded the 90% who need it most.
Third, curriculum personalisation. The Kenyan CBC is fundamentally different from the British curriculum or the American Common Core. It emphasises competency-based assessment, not rote examination. A personalised system must understand which strand, sub-strand, and learning outcome the student is working toward — not just "math" or "science" in the abstract.
Where current approaches fall short: they personalise pace but not context. They say "this student is behind in algebra" but still teach algebra using American examples, American language, and American pedagogical assumptions. That is pace adaptation, not personalisation.
Editorial Commentary: The distinction Aaron introduces is between adapting to a learner and actually understanding one. Most personalised learning systems adjust the difficulty, pace, or sequence of a lesson based on what a student gets right or wrong. Aaron is pointing to a layer deeper, the language a student uses, the examples that make sense to them, the curriculum they are actually following, and even the device and connectivity through which they access the lesson all shape whether learning can happen in the first place. That makes the usual idea of personalisation look surprisingly narrow. A system can know that a learner is struggling with fractions and still misunderstand the learner completely. The code-switching example makes that visible: what the system may interpret as messy input can actually be a perfectly coherent way of thinking and communicating. The challenge, then, is not simply teaching AI more about Africa. It is teaching AI to stop treating the learner's context as noise around the lesson.
Q: MasomoGPT has evolved beyond an AI tutor into a platform that brings together multiple ways of learning. Why was it important to build an integrated learning ecosystem rather than focusing on a single AI product?
A: Because learning does not happen in one mode.
A student might need an AI tutor to explain a concept at 9 PM when no teacher is available. The next morning, she might need a quiz to test whether she actually retained it. Over the weekend, she might want to read deeper into a topic through a digital textbook. Her teacher needs a dashboard to see who is struggling and who is coasting. Her parent needs a progress report to justify the subscription fee.
If you build only the AI tutor, you solve one moment in a complex learning journey. The student still has no assessment loop, no reading material, no teacher oversight. You have a chatbot, not an education system.
We built MasomoGPT as an integrated ecosystem — AI tutoring, adaptive quizzes, an e-book library, AI Classroom Mode for teachers, and a parent portal — because that is what a complete learning experience requires. Each component feeds the others. The quizzes inform the AI about knowledge gaps. The AI tutor addresses those gaps. The teacher dashboard surfaces patterns the AI alone cannot interpret. The e-books provide depth beyond what a conversational AI can deliver in a single response.
The alternative — building one feature and hoping third parties fill the rest — does not work in African markets where those third parties do not exist or cannot integrate with your system
Q: As MasomoGPT expands beyond Kenya into other African markets, it encounters different curricula, languages, educational priorities, and learning cultures. How do you balance building a scalable platform with respecting the diversity of education systems across the continent?
A: Architecture solves this. We built MasomoGPT on a layered system where the AI model is universal but the context layer is local.
The core model — MasomoGpt-v4-PRO, 71 billion parameters — handles reasoning, explanation, and interaction across all markets. It does not change between Kenya and Rwanda. What changes is the MCP (Model Context Protocol) layer that sits between the model and the student. This layer contains curriculum mappings, language hierarchies, assessment frameworks, and pedagogical conventions specific to each country.
Kenya uses CBC with competency-based assessment. Rwanda uses a different competency framework with English and French instruction. Uganda has its own curriculum body. The DRC operates primarily in French with Lingala and Swahili as local languages. Each of these requires a different context configuration — but the same underlying AI architecture.
This means we can enter a new market by configuring the context layer rather than rebuilding the model. Curriculum alignment takes weeks, not years. Language support scales through our multilingual architecture — we already serve seven languages and the system is designed for sixty-one.
But scalability without local partnership is colonialism. In every market we enter, we work with local teachers for curriculum annotation, local linguists for language validation, and local education authorities for regulatory alignment. The technology scales. The relationships do not — and that is by design. A platform that claims to serve Rwandan students without Rwandan teachers in the loop is making the same mistake as every imported EdTech solution: assuming technology alone solves context.
Editorial Commentary: Aaron challenges the assumption that scaling technology means making the technology increasingly indifferent to place. Most global platforms achieve scale by standardising the experience and treating differences between markets as exceptions to be accommodated. Here, the opposite logic is at work in Aaron’s approach. The system becomes scalable because it can carry different contexts without forcing those contexts into the same mould. That distinction matters in education, where curriculum, language, and pedagogy are not surface-level preferences but part of what makes knowledge meaningful to a learner.
It also changes the role of the people involved in each market. Teachers and linguists are not being brought in to help an already finished technology “localise.” They help determine what the system needs to know in order to function there at all. That gives local knowledge a position much closer to the architecture itself. And perhaps that is the more consequential idea in his approach: scale does not have to mean exporting the same intelligence everywhere. It can mean building technology capable of being shaped by the places where it operates. The African continent is already saturated with technologies designed elsewhere, often carrying assumptions about how people communicate, learn, and use technology. MasomoGPT's approach introduces a different possibility: the people and contexts using the technology get a greater say in shaping what the technology becomes.
Q: As AI becomes increasingly capable of explaining concepts, answering questions, and supporting students, educators are rethinking their own role. How do you see the relationship between AI and teachers evolving, and which aspects of education should always remain fundamentally human?
A: I reject the framing of AI versus teachers. It is not a competition. It is a multiplication.
A teacher in Kenya manages 60 students in a class. She cannot individually assess each student's understanding in real time. She cannot provide personalised feedback on 60 homework assignments every night. She cannot be available at 9 PM when a student is stuck on a concept before tomorrow's exam.
AI handles scale. The teacher handles meaning.
What should always remain human: motivation, mentorship, ethical formation, emotional support, and the modelling of how a thoughtful adult engages with knowledge. A student does not learn intellectual courage from an AI. She learns it from watching her teacher wrestle with a difficult question honestly. She does not learn collaboration from a chatbot. She learns it from group work mediated by a human who understands social dynamics.
What AI does better than any teacher at scale: instant feedback, adaptive difficulty, infinite patience for repetition, 24/7 availability, and the ability to meet each student exactly where they are without the cognitive load of managing 59 other students simultaneously.
Our AI Classroom Mode is built on this philosophy. The teacher remains the conductor. The AI is the orchestra — powerful, responsive, but directionless without human leadership. The teacher sets learning objectives, the AI executes personalised delivery, and the dashboard gives the teacher visibility into what is actually happening at the individual student level. The teacher's expertise is elevated, not replaced.
Q: AI opens exciting possibilities for education, but it also raises concerns around trust, misinformation, privacy, and children's interactions with intelligent systems. As someone building AI directly for learners, how do you approach questions of safety, responsibility, and age-appropriate AI use?
A: This is not theoretical for me. I build AI that ten-year-olds use every day. Safety is not a feature — it is a constraint that shapes every architectural decision.
Three principles guide our approach:
First, curriculum boundaries. Our AI is not a general-purpose chatbot. It operates within the bounds of educational content. A student cannot use MasomoGPT to generate harmful content because the model is trained and constrained to respond within curriculum-aligned topics. If a student asks something outside educational scope, the model redirects rather than refuses — "That is an interesting question. Let us explore what your textbook says about..." rather than a blunt "I cannot help with that."
Second, data minimisation. We collect what is necessary for learning personalisation — topic progress, quiz scores, time spent — and nothing else. No microphone access, no camera, no location tracking, no social features where children interact with strangers. The platform is a learning tool, not a social network. All data stays in Kenya on our infrastructure, not routed through US servers.
Third, transparency with parents and teachers. The parent portal shows exactly what the child is studying, what questions they are asking, and how the AI is responding. There is no black box. If a parent wants to review every interaction their child had with the AI this week, they can. Teachers have the same visibility through AI Classroom Mode.
The deeper responsibility question is this: we are training the next generation's relationship with AI systems. If their first experience is manipulative — designed to maximise engagement rather than learning — we are shaping harmful habits. MasomoGPT is deliberately designed to be useful and then get out of the way. We do not optimise for time-on-platform. We optimise for learning outcomes. A session that takes 15 minutes and achieves mastery is better than one that takes 60 minutes and achieves retention.
Editorial Commentary: AI can absorb the repetitive work that makes individual attention difficult at scale, but the parts of education that involve judgement, relationships, motivation, and the formation of a student’s sense of self are being kept deliberately outside that logic. This matters because efficiency is usually the easiest argument for putting AI into a classroom: if a machine can explain, assess, recommend, and respond faster, why not let it do more? Aaron’s answer is essentially that being able to automate something is not the same as having a good reason to automate it.
There is another implication in his decision not to optimise for time spent on the platform. Most digital products are designed around keeping people engaged for longer. An educational system designed around mastery has a very different incentive: the best outcome may be for the student to leave the platform sooner because they no longer need it. That is a small but significant rejection of the attention economy, particularly when the users are children. It suggests that responsible educational AI is not only about preventing harmful outputs; it is also about deciding what kind of relationship we want young people to develop with technology in the first place.
Q: Technology companies often measure success through growth and engagement, while education is ultimately about meaningful learning. As MasomoGPT continues to grow, how do you evaluate whether the platform is genuinely improving educational outcomes rather than simply increasing usage?
A: We track three things that matter and deliberately ignore the vanity metrics that most EdTech companies optimise for.
First, learning gain. We measure pre-test to post-test improvement on curriculum-specific assessments. Not "time on platform" or "sessions per week" — actual knowledge gain as measured by CBC or Cambridge-aligned assessments. A student who uses MasomoGPT for 10 minutes and demonstrates mastery is a better outcome than a student who scrolls for 60 minutes and learns nothing.
Second, pilot-to-paid conversion. This is our proxy for genuine value. If a school tries MasomoGPT for free and then chooses to pay for it, that is the market telling us the platform works. Our current conversion rate is 68%. Schools do not pay for tools that do not work — especially in price-sensitive African markets where every shilling is scrutinised. That number is the strongest signal we have that the platform delivers real educational value.
Third, teacher validation. We ask teachers: "Did your students perform better on curriculum assessments after using MasomoGPT?" Not our assessments — their assessments, designed by them, scored by them. If the teacher says yes, and the test scores confirm it, then the platform is working. If usage is high but teacher-reported outcomes are flat, something is wrong regardless of what our engagement metrics say.
What we deliberately do not optimise for: daily active users, session length, notification-driven re-engagement, streaks, or gamification loops. These are addiction mechanics dressed as engagement. Education is not entertainment. The goal is competence, not compulsion.
Editorial Commentary: There is an almost counterintuitive measure of success hiding in MasomoGPT's philosophy: the platform should help students become more capable, not more dependent on the platform itself. A student who gradually needs less support from the system while performing better at school could therefore represent a stronger outcome than one who simply spends more time using it. Measuring that relationship will require more than comparing usage with grades. Learning is shaped by teachers, schools, home environments and many other factors, so the real opportunity for Aaron is to build increasingly rigorous ways of showing where its contribution sits within that picture. That could mean tracking learning gains over time, comparing patterns across classrooms, following students beyond individual sessions, and combining platform data with teacher assessments. The challenge is considerable, but so is the opportunity: if MasomoGPT can demonstrate that its intervention is associated with growing student capability and independence, it moves beyond proving that students use the platform to proving that the platform is actually changing what students can do. That is the kind of disruption that turns technology from something students continually return to into something that gradually makes them more capable without it.
Q: You've argued that education should become "an engine for wealth creation, not just wealth consumption." What does that philosophy mean in practice, and how should AI reshape the connection between learning, skills, and economic opportunity?
A: The traditional model is: pay for education fi get a certificate fi hope someone hires you fi start earning. Education is a cost centre for the first 16-22 years of a person's life. That is wealth consumption.
The model I believe in: learn a skill fi apply it immediately fi generate income fi reinvest in deeper learning. That is wealth creation.
In practice, this means MasomoGPT does not just teach curriculum content — it builds capability that has immediate economic value. A student who learns data analysis through our platform can freelance on Upwork next month. A student who masters financial literacy can manage her family's M-Pesa business more effectively today, not after graduation.
AI reshapes this connection by collapsing the time between learning and application. Traditional education batches knowledge delivery into years-long programmes. AI enables just-in-time learning — the student learns what she needs, when she needs it, and applies it immediately. The feedback loop is hours, not years.
For Africa specifically, where 60% of the population is under 25 and youth unemployment exceeds 30% in most markets, we cannot afford a model where learning is disconnected from earning for two decades. Education must generate economic value while it is happening, not only after it concludes.
Q: Around the world, education systems are being challenged to prepare learners for an AI-driven economy. What do you think African education systems have an opportunity to do differently, rather than simply following models developed elsewhere?
A: Africa has the youngest population on Earth. Sixty percent under twenty-five. We are not retrofitting an existing system — we are building for a generation that will learn primarily through AI interfaces.
That is an opportunity no other continent has at this scale. Three things we can do differently:
First, skip the textbook era entirely. Europe and America digitised textbooks into PDFs and called it EdTech. Africa can leap directly to adaptive, AI-native learning that has never been bound by a static textbook. Our students' first serious educational technology interaction can be intelligent from day one — not a digital photocopy of a printed page.
Second, build multilingual AI-native education from the start. The rest of the world is trying to retrofit multilingual support onto systems built in English. We are building multilingual from the ground up. Our model handles code-switching natively because our students code-switch natively. This is not a limitation we are accommodating — it is a feature we are designing for. In ten years, the world will look to African language AI as the gold standard for multilingual education systems because we had no choice but to solve it properly.
Third, design for economic agency, not institutional dependence. Western education systems were built to produce employees for industrial economies. Africa can build education systems that produce entrepreneurs, creators, and self-directed learners for a post-industrial economy. AI makes this possible because it can personalise learning paths around each student's economic context and opportunity set — not funnel everyone through the same standardised pipeline toward the same credential.
The worst thing Africa can do is wait for OECD countries to figure out AI in education and then copy their homework. By the time they have a framework, we should have a functioning system.
Q: If your vision succeeds over the next decade, how would you hope a student using AI-enabled education in Africa experiences learning differently from students today?
A: I want a student in Turkana to have the same quality of personalised education as a student at a private school in Nairobi — not because we lowered the ceiling, but because we raised the floor.
Specifically, in ten years, I hope that student experiences this:
She opens her phone — whatever device she can afford — and her AI tutor already knows what she struggled with yesterday. It presents today's lesson in her mother tongue, code-switches into English for technical terminology, and adjusts difficulty in real time based on her responses. If she is stuck, it explains differently — not louder. If she masters something quickly, it moves on without making her sit through material she does not need.
Her teacher sees a dashboard that morning showing exactly which students need intervention and on which topics. The teacher spends her time on the students who need human connection, not on repetitive explanation that the AI handles better at scale.
Her parent receives a weekly message — in their language — showing what their child learned, where they are strong, and one specific thing they can do at home to support the next week's learning.
The student never encounters a paywall that locks her out of knowledge her richer peers can access. The economics work because the AI runs on infrastructure we built and own — not rented per-token from a company in San Francisco.
Most importantly: she experiences learning as something that responds to her. Not a system she must conform to, but a system that conforms to her. That shift — from student-serves-system to system-serves-student — is the transformation I am building toward.
Q: What is something about AI and education that is widely discussed but still fundamentally misunderstood?
A: The cost question.
Everyone discusses whether AI will "replace teachers" or "personalise learning" or "democratise access." These are important conversations. But the thing that is fundamentally misunderstood — the thing that will determine which African EdTech companies survive and which collapse — is the economics of AI inference at scale.
Here is the math that nobody in the "AI for education" conversation wants to confront: a single student session burns approximately 2,000 tokens. At foreign API rates, that is $0.012 per session. Twenty sessions per month across 100,000 students costs $24,000 per month in inference alone — before you build anything, before you pay anyone, before M-Pesa takes its cut. And your revenue ceiling in the mass African market is 500 Kenyan shillings per month — about $4. The unit economics do not work if you rent your intelligence from someone else. This is why we trained our own model. MasomoGpt-v4-PRO serves inference at under $0.0001 per query — two orders of magnitude cheaper than any foreign API. At ten million students, our inference costs approximately $50,000 per month. The same ten million students on a foreign API would cost $2.4 million per month. That is not a competitive advantage. That is the difference between a viable business and an impossible one. The widely discussed version of this conversation is "should Africa build its own AI?" — framed as a philosophical or sovereignty question. The actual answer is not philosophical. It is arithmetic. If you want AI education to reach the students who need it most, at a price point they can afford, you must own the model. Full stop. Everything else is a subsidy waiting to expire.
Editorial Commentary: The most consequential part of Aaron’s argument may be hiding in the least glamorous part of this section: the cost of running the intelligence. We tend to discuss African-native AI in terms of language, culture, representation and sovereignty, but there is a more basic question underneath all of that: who can afford to run the system at scale? If every interaction with an AI tutor creates a bill to an external model provider, then the ambition to make high-quality personalised learning affordable to millions of African students is constrained before the product even reaches the classroom. The question of whether to build or rent the underlying infrastructure therefore becomes an economic question, not just a technological or nationalist one.
That connects directly to his broader idea of education as wealth creation. If the technology itself carries a cost structure designed for wealthier markets, then the students who most need affordable, continuous learning remain the hardest to serve sustainably. Owning more of the underlying infrastructure changes that equation. It gives Aaron room to design around the economics of African households rather than trying to squeeze African education into the pricing model of technology built elsewhere. And that may be the more interesting meaning of “African-native” here: not simply technology that speaks African languages, but technology whose economics make African-scale participation possible.
That is a much bigger proposition than building an African AI tutor. It is an attempt to make the cost structure of the tool part of the design of the education system itself.
What makes this conversation with Aaron particularly interesting is how far it moves beyond the familiar promise that AI can make education more personalised. His ambition is not simply to put an AI tutor in front of more African students, but to rethink the relationship between the learner, the technology, and the system around them. From code-switching and local curricula to low-cost devices, teacher involvement, infrastructure ownership, and the economics of AI inference, his answers keep returning to the same underlying question: what would education look like if the system were designed around the realities of the learner rather than asking the learner to adapt to the system?
That question becomes particularly important in Africa, where education technology is often introduced through products and architectures designed elsewhere. MasomoGPT represents an attempt to approach the problem from the other direction: build around African learners first, and let the technology emerge from those realities. The ambition is not only to make AI more accessible, but to make its underlying economics, intelligence, and learning experience capable of working at African scale.
There is still a great deal to prove. Building an education platform that can demonstrate measurable learning gains across different countries, languages, curricula, and economic contexts is a significant undertaking. But perhaps that is precisely what makes the journey worth watching. When MasomoGPT succeeds, its significance will extend beyond the students using the platform. It could offer a different answer to what it means to build technology for Africa: not adapting an existing model to a new market, but building the infrastructure, economics, and intelligence around the realities of the people the technology is meant to serve.
And perhaps the most compelling measure of that success will not be how many students MasomoGPT keeps on its platform, but how much more those students are eventually empowered and become capable of doing without needing the platform.
Thank you for reading!
Don’t see your pick in the options? Drop it in the comments. Aaron joins the map this weekend.
Thank you for standing with this work.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.