Gradely AI Tutor

Inspiration

Nine out of ten children in Sub-Saharan Africa cannot read and understand a simple text by the age of ten. Not because they are out of school. Most of them are sitting in a classroom of sixty, where one teacher cannot possibly work out what each child misunderstood on Tuesday and fix it by Thursday. Families who can afford a private tutor buy their way out of the problem. Everyone else carries the gap forward until an exam makes it official. We have been building for those children since 2021 and have reached over 200,000 of them. But there was an honest limit to what the platform could do. It could tell a parent exactly what their child got wrong. It could not sit down with the child and teach. That is the gap this project closes. One-to-one attention is the scarcest resource in African education, and it is the one thing a model like Gemini can actually manufacture.

What it does

Gradely AI Tutor teaches back in real time. A child works through a problem and the tutor follows how they are reasoning, not just whether the final answer is right. When it spots the misconception it does not hand over the answer. It asks the next question that gets the child there themselves. Three things separate it from a chatbot with a syllabus attached: It answers with video. When a gap appears mid-session, the tutor generates a personalised explanation video for that specific concept, on the spot, rather than pointing the child at a library. It reacts like a character. Instant avatar feedback on every answer, coins, badges, streaks and leaderboards. The tutor children keep opening is the one that feels alive. It hands off. When the AI is not enough, the child reaches a human tutor inside the same chat, and that tutor arrives already knowing the diagnosis, the misconception history and what has been tried. Every session is grounded in the child's grade, their learning history and our library of over 100,000 curriculum-aligned items and 3,500 animated lessons, so the tutor teaches the curriculum the child is actually sitting rather than a generic approximation of it.

How we built it

Challenges we ran into

Gemini is the intelligence and orchestration layer. Every tutoring turn passes it the tutor persona and behavioural rules, the student's grade and learning history, the retrieved curriculum content relevant to the moment, and the live dialogue. Gemini decides what happens next in the lesson. Around that, sits the existing Gradely platform: Go and Gin microservices, MySQL and Redis, a vector retrieval layer over the curriculum corpus so the tutor is grounded in verified material rather than model recall, WebSocket streaming so responses and video generation progress arrive live, Firebase and AWS S3 for delivery, and Paystack for payments. Teaching without telling. The hardest part was not getting good answers out of the model, it was stopping it from giving them. Scaffolding and Socratic questioning had to be enforced as behavioural rules, because the helpful default of any LLM is to solve the problem for you. Grounding. A tutor that invents a plausible-sounding method is worse than no tutor. Retrieval had to keep every explanation tied to the curriculum the child is examined on, across both Nigerian and international syllabuses. Latency and bandwidth. Our users are on low-cost devices and thin connections. Streaming responses and generating video had to work in that environment. Cost per session. At telco pricing of around ten cents a day, the unit economics only work if inference and video generation are tightly controlled. Every architectural decision had a cost ceiling attached to it.

Accomplishments that we're proud of

Gemini runs the entire tutoring loop in production, serving real users, rather than sitting beside the product as a demonstration call.
The tutor answers with a generated video explanation in-session, which is the feature we believe no one else is shipping to this market.
Measured learning gains: 46.7% of engaged students improved immediately, at an average of 7.9 percentage points, and students entering at a 10 to 30 percent baseline improved by 20 to 40 percentage points.
Users who stay: NPS of 95, weekly CSAT of 4.78 out of 5, and a 95% class completion rate in premium cohorts.
Reach that makes the impact real: over 200,000 students across 8 countries, distributed through telco partnerships from as little as ten cents a day.

What we learned

Diagnosis has to come before instruction. Most tutoring, human or automated, starts by repeating what the child is already being taught. Starting from the specific misconception is what produces the gains.
Children stay for the character, not the content. The same explanation lands differently depending on whether something on screen reacted to them first.
Grounding beats raw capability. A smaller, well-grounded, curriculum-aware response teaches better than a more impressive ungrounded one.
AI changes the shape of the business, not just the product. Gross margin moved from 45% to 65% in a year as AI absorbed work people used to do, which is what makes ten-cents-a-day pricing survivable.

What's next for Gradely AI Tutor

The build order is set. Next come richer question types, which means voice input, handwriting, essays, on-screen formulas and interactive exercises, then audio out and video in so the tutor speaks and can see and hear the child back.
After that, deeper gamification with a game world and storyline, and a tutor marketplace where the coins children earn buy live human help inside the chat.
The longer bet is a relational voice that learns how to talk to this particular child, metacognitive check-ins that ask by voice how they are thinking about a problem, and an avatar that follows the child across the apps they already use.
Three experiments are queued to test whether this actually works: whether a human tutor rates the AI's diagnosis as correct when they diagnose the same child independently, whether the gamified cohort returns at 30 days more than a control, and how many days it takes to close a gap against a control group.
Commercially, the tutor rolls out through the existing telco channels in Nigeria and into the UK diaspora launch.