Africa is short 15 million teachers.Nobody can hire that many.
Students worldwide can teach — but a twenty-year-old can also teach an error. We verify every correction by symbolic computation, which makes an abundant, non-expert tutor as reliable as a scarce, expert one.
- 15 M
- teachers missing by 2030
- UNESCO, Sub-Saharan Africa
- +170 M
- more school-age children by 2050
- 280 M in 2020 → 450 M
- $15–40
- per hour, private tutor in Nairobi
- IGCSE maths and sciences
- $60–160
- a month for four hours
- the budget we replace
The software, running
Pick a paper. Watch it get diagnosed.
Real rows from our evaluation harness — nothing staged. The last case is one the engine cannot explain, and says so.
Derivative(x*exp(2*x), x)2*exp(2*x)Line 2 no longer follows. Decided by computation, not by a model.
Cause
Derivative of a product computed as the product of the derivatives
product_rule_as_product_of_derivatives
First question to ask
Ask them to differentiate x·x with their method, then compare to the derivative of x². The counter-example breaks the rule better than restating the correct one.
Verified: this rule regenerates line 2 exactly.
Errors located — deterministic, no model
Reproduced by a named rule
Once the library extends itself
What these numbers are worth. The corpus was written alongside the rules — it shows the mechanism works, not that it generalises. Validation on 50 unseen papers is the current milestone. We publish the limit, not just the figure.
The obvious objection
« Isn't this a wrapper around a language model? »
Partly, and we would rather say so. A frontier model reads the handwriting and writes the prose — two stages out of six. Given a careful prompt it will often name the misconception too. We do not claim otherwise.
What a bigger model does not fix is the verification. Checking an answer requires an oracle independent of whatever produced it. If the same system both writes the correction and grades it, a systematic error passes twice — and the student, who cannot referee the maths, learns it. Our checker is a computer algebra system. It has no opinion about whether line 3 follows from line 2; it computes it. That is a property of the architecture, not of model capability, and no amount of added context supplies one.
The rest follows from the same test: does it live inside one call, or between them? A better explanation lives inside the call and will be overtaken. A library of named wrong rules, a student model that remembers an error from three weeks ago, a proof that a correction is sound, and paper on a desk all live between calls — and someone has to build and maintain them. That is the company.
We are running the honest version of this test: the same corpus against a single general-purpose model call, published whichever way it lands. If the model matches the library, the library's value is auditability, cost and accumulation rather than accuracy — and we would rather find that ourselves than in someone's diligence.
The device
It stands at the head of the notebook
Nothing to hold, nothing to launch, no phone of the student's to depend on. Camera, light and screen in one arch that watches the page continuously — so an error is caught the second it is written rather than ten minutes later. A tutor can join through that screen at any moment, without the student having to ask.
- 150 mm
- Height under the crossbar
- Low enough to glance at without lifting your head from the page
- 50 mm
- Setback behind the page
- What makes the camera angle achievable with a standard lens at all
- 42°
- Optical axis from vertical
- A 48° vertical field covers the full 297 mm depth of an A4 page
- 8 MP
- Sensor floor
- Below it the bottom of the page drops to 138 dpi and exponents stop resolving
Everything in one arch
A tilted camera and a diffused light bar under the crossbar — without that light, a desk lamp throws the hand's shadow across the writing and one capture in two is lost. A four-inch screen on top, a speaker and a microphone inside. On the legs, a camera-live indicator wired to the sensor's own power and a mechanical switch that cuts it.
About $73 in parts
More than the $30 we first aimed at. The screen, the audio and the 8 MP sensor are $41 of it — that is the price of the tutor mode and of an object that needs nothing else on the desk. If the budget forces a cut, the screen goes before the camera.
The business
We start where we can get paid, to get where we have to go
Our first customers are not the families in the 15-million-teacher statistic. They are English-medium IB and IGCSE households in Nairobi, Dar es Salaam, Kigali, Lagos and Accra — who pay in hard currency, share a time zone with European tutors, and already buy private tutoring at $15–40 an hour. That budget funds the descent down-market; it is the entry point, not the destination. Every device sold amortises the engine, and every hour taught trains it — until the marginal cost of a good correction approaches zero and families who could never pay become servable.
Why this could not have been built in 2015
The wave is arriving, not forecast
The children who make Sub-Saharan Africa go from 280 to 450 million school-age kids are already born. The teacher shortfall is dated 2030, not 2050.
The object finally costs $73
An 8 MP camera module and a $15 computer put a device that watches paper inside a family budget. In 2015 the same build was an order of magnitude more.
Models only just learned to read handwriting
Reliable transcription of messy handwritten maths is roughly two years old. It is the one stage we could not have built ourselves at any price.
Why hardware does not sink us
A $75 device on a $12 subscription takes eight months to repay itself — which is not a business. So we price a learner, not a box. The camera does not care how many notebooks pass under it. Siblings and study groups share one arch, and the payback that decides whether this is fundable moves from eight months to under three.
Sharing is not a loophole in the model — it is the model, and it is what makes the access real rather than rhetorical.
Priced per device
1 student, 1 arch
8.1 months
$9.29 contribution / month
Too slow to finance a fleet.
Priced per learner
3 students sharing 1 arch
2.6 months
$28.87 contribution / month
At $12 per student per month, against the $60–160 a family already spends on four hours of tutoring.
These are modelled, not measured. The hourly rate they rest on comes from desk research; replacing it with numbers from live interviews with Nairobi families is work in progress, and it is the input that moves this table most.
What compounds
The camera is the wedge. The record is the company.
The record
Measures the work, not just the answer
Hesitation before a step. Lines crossed out and rewritten. The hour at which accuracy collapses. Every assessment ever built records what a student produced; this records how — and that data exists at scale for no population anywhere.
The compounding
Each hour observed is an hour nobody pays for
The library of named errors grows from every paper it cannot yet explain. As it covers more, the engine handles more alone, and human hours move to what software will not do — motivation and accountability.
On holding a record of a child
Video is processed on the device and never leaves it — only the line that changed is analysed. A hardware indicator shows when the camera is live and a physical switch cuts its power. Kenya's data protection law applies to us from the first unit, and a model trained on African students belongs in African schools.
Who is building this
Clément Duméril
Mechanical engineering, Arts et Métiers (ENSAM) · MEng, UC Berkeley
I sit on both sides of this problem. The supply side of the network — European engineering students who can teach STEM — is the one I come from and can recruit directly. The hardware is what I was trained to design. And this is not a plan: the diagnostic engine below is written, running and measured.
Stated plainly rather than left to be found: I am a solo founder, working on this part-time until my MEng finishes in May 2026, and full-time from then. Burn to date is about zero. I am looking for a technical co-founder at Berkeley.