I am honestly excited to share SciNova AI, an AI-powered learning ecosystem I am developing to help secondary school students learn STEM subjects more effectively, especially in situations where access to high-quality science education, personalised support, laboratories, and experienced teachers is somewhat limited, or just not really there.
SciNova AI is not just an AI chatbot for students or something like that, it’s more like a whole ecosystem.
The larger vision is to build a Personalized Intelligence Learning Ecosystem for secondary STEM education, one that can spot what a learner understands, surface misunderstandings and learning gaps , adjust instruction to individual needs, offer chances for practice and experimentation, and then give teachers useful information about how their students are learning , you know not in a vague way but in a clear way.
I am developing SciNova AI from Nigeria, with secondary STEM education in mind first, but I might end up serving learners across other low-resource educational contexts too later on.
More importantly, I am approaching the project through an Effective Altruism lens: I’m not assuming that an exciting technology is automatically impactful. Instead I keep asking whether there is a believable route where it could create real, substantial benefits, whether those benefits can be measured, whether the problem is sufficiently neglected, and whether the intervention can be made cost-effective and scalable.
The problem I am trying to solve
My experience as a science teacher in under-resourced communities in Lagos State Nigeria keeps showing me a problem that people often underestimate, and it’s easy to miss.
Students do not only differ in “knowing” and “not knowing” a topic. They differ in what they actually understand, why they are stuck, how fast they learn, what misconceptions they carry, and what kind of explanation or real experience helps them move forward.
Still, a lot of secondary education works at the level of the class, not the individual learner.
A teacher can have 40, 60, sometimes even more students in one classroom. So it becomes extremely hard to continually determine things like:
- Which concepts each student has genuinely understood
- Which misconceptions are blocking further learning
- Which students are falling behind quietly , without making noise
- What type of explanation might work best for one particular learner
- Whether extra practice is actually fixing the root issue
- Which students need more challenging material, not just repeated review
- How learning difficulties are spread across the class
This problem becomes especially huge in STEM subjects, because foundational misunderstandings can linger on for long and affect the understanding of other related concepts.
For example, if a student misunderstands cell division, then they may later struggle with genetics, reproduction, and other biology areas. Just giving that student more questions may not fix anything if the initial misconception isn’t identified.
At the same time, many schools in low- and middle-income contexts face constraints in teacher availability, laboratory access, learning resources, and truly individualised support.
So here is the question that keeps coming up:
Can AI substantially increase the amount of high-quality, personalised STEM learning support available to students without needing a proportional increase in educational resources?
I don’t yet know the answer.
SciNova AI is basically my attempt to investigate it.
What is SciNova AI?
SciNova AI is envisioned as an AI-powered conversational learning ecosystem for secondary STEM education.
The initial target is students around the JSS–SSS / KS3–KS5 range, with support for curricula and examinations like WAEC, NECO, UTME, GCSE, IGCSE, and A-levels.
The platform is being designed around a set of connected abilities, a bit like a chain where one part helps the other, even if that link is not always obvious.
1. Personalised AI tutoring
Students can talk with an AI tutor conversationally, not just get static pages of content and move on.
The point is to go beyond "Here is the answer."
towards "What does this student currently understand, where is their reasoning going wrong, and what intervention is most likely to help?"
That also means the AI tutor needs to be sensitive to what the learner has already done, said, or implied earlier, instead of treating each question like it’s a stand alone chat, like nothing else matters.
2. Misconception detection
One area I’m really drawn to is diagnostic learning.
Instead of only logging whether someone got a question right or wrong, SciNova AI is being developed to find patterns in incorrect answers and in the reasoning that sits behind them, which might point to underlying misconceptions.
For instance, two students might both miss the same physics question, but they could be missing totally different things, so the “fix” should be different too.
So the system should help tell apart:
"Student got the question wrong."
from:
"Student appears to hold misconception X, which is likely contributing to this pattern of errors."
That kind of split could make personalised learning a lot more useful, because the intervention actually matches what’s happening.
3. Adaptive learning
Once the system has evidence about a student’s understanding, it can adapt things like:
- explanations
- examples
- questions
- difficulty
- learning sequence
- revision activities
- feedback
- remediation
The longer-term idea is to build a learning system that gets more responsive over time, kind of more attuned to the individual learner, not just initially customised once.
4. AI-supported teacher intelligence
SciNova AI is not meant to replace teachers.
A big piece is a dashboard that teachers can use.
Teachers could potentially see things such as:
- common misconceptions across a class
- individual learning gaps
- student progress
- engagement patterns
- difficult concepts
- areas requiring intervention
- learning trajectories
In other words, AI could act like a force multiplier for teachers, rather than being only a student-facing chatbot that they mostly ignore.
5. Virtual STEM laboratory
Another part of the longer-term vision is a SciNova Africa Virtual STEM Lab / Digital Science Twin.
This is especially relevant in settings where real lab resources are scarce or just not reliably available.
The goal is not to argue that simulations can simply swap out physical laboratory experiences. Instead, virtual environments could give students extra chances to explore scientific phenomena, practise procedures, and build conceptual understanding, particularly when physical resources are missing, unavailable, or insufficient.
Why approach this through Effective Altruism?
My interest in Effective Altruism has shifted the way I think about educational innovation, a lot.
There are countless educational problems, and also countless potential interventions. Simply building something technically impressive is not enough.
The more important questions are kind of like:
How big is the problem?
How much could this intervention really improve outcomes?
For whom exactly?
Compared with what baseline?
How neglected is this opportunity?
How tractable is the intervention in practice?
Can its effects be measured properly?
Can it reach people at low enough cost?
What risks and unintended consequences could show up?
These questions matter especially for AI in education, because the technology can create both upside and downside, sometimes at the same time.
SciNova AI is therefore being developed as an empirical hypothesis, rather than a predetermined “answer” that I’m already assuming is true.
The hypothesis, roughly, is:
Highly personalised, AI-supported STEM learning could substantially improve learning outcomes for students who currently have limited access to individualised educational support, and could potentially do so at a sufficiently low marginal cost to make large-scale deployment feasible.
That’s a claim I want to test, not a conclusion I’m leaning on.
Where the EA principles come in
There are a few ways this project could connect to EA-style prioritisation.
Scale
There are huge numbers of students globally who need better educational support.
STEM education is especially important because scientific and quantitative literacy can shape students’ future education and career paths, while also feeding broader human capital.
If a scalable intervention creates meaningful improvements in learning outcomes, the total benefits could be very large in the aggregate.
That said, scale by itself is not proof of impact. A massive group with a tiny effect might end up less valuable than a smaller group with a very large effect.
That’s one reason I want rigorous evaluation to be central to SciNova AI.
Neglectedness
AI education has already gotten substantial attention worldwide, so I wouldn’t claim that “AI in education” overall is neglected.
But the opportunity that might be more interesting is narrower.
There may be substantial room to develop and evaluate AI-powered, personalised STEM learning systems aimed at students in under-resourced educational contexts, not mainly optimising for well-resourced markets where the incentives already exist.
My own starting point is Nigeria and potentially other African contexts.
This gives the project an opportunity to look at questions that may receive less attention than AI tutoring for affluent educational markets.
I would like evidence, not assumptions, to figure out whether it is genuinely neglected , or if that idea just sounds nice.
Tractability
AI systems are becoming more and more capable of giving personalized explanations, making questions, analysing learner responses, and adapting interactions.
So, the technical side of personalized learning is getting more tractable.
But the hard piece is whether those capabilities actually turn into measurable improvements in real learning, not only in demos.
A system that students enjoy using, but does not improve learning would basically weaken strong impact claims.
Cost-effectiveness
One potentially interesting property of software-based AI interventions is the possibility of very low marginal costs after development.
If an AI learning system eventually showed a meaningful effect on learning, and could be deployed at relatively low cost, then it might have real potential for scale.
But again, this is conditional.
Development, inference, infrastructure, teacher training, device access, connectivity, monitoring, and safeguarding all come with costs.
So I see cost-effectiveness as something to measure, not something to assume.
What I have built so far
SciNova AI has evolved a lot, from the original concept.
At first, I pictured a fairly straightforward AI tutor.
But through experimentation and reflection, the concept has drifted into something bigger: a broader Personalized Intelligence Learning Ecosystem.
The current architecture includes or is being built around:
- AI conversational tutoring
- persistent learner context
- diagnostic learning
- misconception detection
- adaptive quizzes
- gamified learning
- student progress tracking
- teacher dashboards
- learning analytics
- curriculum-aligned STEM content
- virtual laboratory experiences
- multiple communication channels
- future AI-powered science simulations
Right now the system is being developed as a prototype, and pilot testing is ongoing rather than it being presented as a validated educational intervention.
That difference matters.
I have built a technology prototype. I have not yet shown the causal educational impact of the complete system.
What I don't know yet
There are several key uncertainties.
Will AI tutoring actually improve learning outcomes?
Better explanations do not necessarily equal better learning.
Students might misunderstand AI-generated explanations , or become overly dependent on the system, or they might just interact with it without really engaging with the underlying concepts.
Can AI reliably diagnose misconceptions?
This is one of the most technically and educationally interesting questions.
A wrong answer does not automatically reveal a misconception.
We need better methods for separating out:
- random errors
- knowledge gaps
- misconceptions
- language difficulties
- careless mistakes
- misunderstanding of the question
- deeper conceptual models
Will students actually use the system?
Adoption and sustained engagement are big uncertainties here.
And will teachers trust , and then actually use, the information?
Teacher dashboards only matter if the information is accurate, interpretable, and genuinely useful for instructional decisions.
Could AI introduce new educational harms?
Potential risks include:
- inaccurate explanations
- hallucinated information
- reinforcing misconceptions
- excessive dependence on AI
- reduced student agency
- privacy concerns
- inappropriate collection of learner data
- algorithmic bias
- inequitable access
- weakening rather than strengthening teacher–student relationships
These risks need to be treated as part of the research problem, not just as an afterthought.
What I want to do next
My immediate priority is not simply to add more features.
It is to establish whether the core intervention works, first.
I want to move through increasingly rigorous stages, kind of like:
Stage 1: Technical validation
Determine whether the diagnostic and tutoring architecture works reliably.
Can the system identify meaningful learning patterns?
Can it keep an appropriate learner context?
Can it generate useful interventions?
Stage 2: Usability testing**
Work with a small number of students and teachers to identify:
- usability problems
- confusing interactions
- inappropriate explanations
- workflow issues
- teacher needs
- student engagement patterns
Stage 3: Educational pilot
Compare learning outcomes before and after using the system, while collecting qualitative evidence about how students and teachers experience it.
Stage 4: Controlled evaluation**
If the early evidence is promising, conduct a more rigorous evaluation that can estimate the causal effect of the intervention.
Ideally it would include appropriate comparison groups and pre-registered outcome measures.
Stage 5: Cost-effectiveness and scalability
If educational effects are shown, investigate whether the intervention can deliver those benefits at a cost that makes large-scale deployment worthwhile.
Why i am sharing this on the EA Forum
I am posting SciNova AI here because I want the project to get seen by people who can ask uncomfortable but useful questions.
I am especially looking for feedback on stuff like,
- if the problem is actually important enough
- whether the theory of change i am proposing is even plausible
- whether there are stronger intervention opportunities i should directly compare against
- how to design an informative early-stage evaluation
- what educational outcome measures make sense (not just “nice to have”)
- how to estimate cost effectiveness without fooling myself
- potential AI safety and misuse risks , and what failure modes might be “quiet” but serious
- whether this area is genuinely neglected compared to other opportunities
- how to avoid building something technically impressive but low impact, like pretty demo syndrome
- possible collaboration opportunities with researchers, educators, and EA aligned orgs, if that fits
I’d really value criticism that could make me change things, or abandon parts of the project entirely.
If the evidence eventually suggests that SciNova AI is not an effective way to improve STEM education, I want to know that.
And if the evidence suggests that a different intervention would produce substantially greater impact, I want to know that too.
From building technology to testing impact
One of the biggest lessons I’ve taken from Effective Altruism is that good intentions are not enough
As an educator i naturally want to build tools that help students.
As a technologist, I am also excited by what AI makes possible, like the whole space feels too big not to try.
But neither of those things on their own establishes that SciNova AI is actually valuable.
The real test is whether it can produce measurable improvements in learning and educational opportunity, especially for students who are currently underserved, and whether those improvements justify the resources needed to achieve them.
So i’m deliberately treating SciNova AI as a hypothesis to investigate, not as a fait accompli.
The ambition is large:
AI-powered science learning that can give every student access to personalised, high-quality STEM support, regardless of where they live or the resources available in their school.»
But the standard of evidence should be, like, equally ambitious.
Build. Measure. Compare. Learn. Iterate.
And if the evidence says we should change direction, well then, change direction.
That’s the approach i hope will carry SciNova AI from an interesting EdTech prototype into a potentially evidence backed intervention for STEM education.
I would be grateful for your questions, criticisms, research suggestions, and potential collaboration, if there’s overlap.
Kehinde Olumuyiwa Adesina
Science Educator | Digital Science Education & AI Pedagogy Researcher | AI & Pedagogical Innovator
Nigeria
