Multiplying teaching capacity

Africa is short 15 million teachers.Nobody can hire that many.

So the capacity has to come from somewhere else. Students worldwide can teach — but a twenty-year-old can also teach an error. Walimu verifies every correction by symbolic computation, which is what makes an abundant, non-expert tutor as reliable as a scarce, expert one.

Engine running today · see it diagnose a real paper

Why the gap bites

Scarce teachers make private help expensive

When qualified STEM teaching is rare, the families who can pay bid for it and everyone else goes without. That is the mechanism we attack — not school fees.

Nairobi, IGCSE tutor

KSh 2,000–5,000

per hour · roughly $15–40

Specialist maths and physics tutors sit at the top of that range and above.

General maths tuition, Kenya

KSh 800–1,500

per session

Rates run higher in Nairobi than elsewhere, and higher again for sciences.

Four hours a month with an IGCSE maths tutor in Nairobi costs a family roughly $60 to $160. We are building for a small fraction of that, priced per student rather than per device so a household or a study group shares one setup.

Rates from published Kenyan tutoring-market sources, 2026. We are validating them with direct family interviews before we set our own price.

Why not just a chatbot

Three things a general model does not do

01

It cannot tell you it is wrong

Misread an exponent, slip on a sign, and the student has no way to detect it. They learn the error. A tool that is confidently wrong to a reader who cannot check is worse than no tool.

02

It corrects after the fact

You photograph a finished page, so the student has already built ten minutes of work on a dead line. The correction arrives when the reasoning is cold and the mistake is buried.

03

Nobody is expecting them

A chatbot will explain the same thing a hundred times and never notice the student stopped showing up. Teenagers do not primarily lack explanations — they lack accountability.

The software, running

Pick a paper. Watch it get diagnosed.

These are real rows exported from our evaluation harness — the same code that produced the numbers below. Nothing is staged. The last case is one the engine cannot explain, and says so.

What the student wroteIB Math AA HL
1Derivative(x*exp(2*x), x)
22*exp(2*x)

First line that no longer follows: line 2. Decided by symbolic computation — no model involved.

What the tutor receives

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.

100%21 of 21

Errors located — deterministic, no model

90.5%19 of 21

Reproduced by a named rule

95.2%after one pass

Once the library extends itself

What these numbers are worth. The starting corpus was written alongside the rules, so it shows the mechanism works — not that it generalises. Validation on 50 unseen student papers is the current milestone. We would rather publish that limit than the figure alone.

Hardware roadmap · not yet built

A 15-gram clip that instruments any pen

The camera above the page reads what is written. The clip reads how it gets written — stroke order, hesitations, the crossing-out. No existing education dataset contains that signal.

Target bill of materials around $30 at prototype volume, for roughly 15 grams — the same order as a smartpen that already ships.

Sensor clip — schematicrev. A
ANY STANDARD PEN01 OPTICAL FLOW02 IMU03 BLE04 BATTERY
PMW3901

Optical flow sensor

Measures real movement across the paper

This is the optical-mouse principle: the sensor watches the paper fibre travel underneath and derives motion directly, at high rate and with no printed pattern. It is what removes the drift that makes an accelerometer alone useless for handwriting.

Why the clip does not replace the camera

In mathematics, position carries meaning: x² and x2 differ only by a height, a fraction is vertical, a matrix is two-dimensional. The clip measures the gesture but ignores the layout; the camera captures the layout but ignores the gesture.

The two are fused, not competing. And the camera already works: it is a phone on a stand, with nothing to manufacture.

Where this stands. The spec is costed and every component is off the shelf, but the clip is not built — count three to six months to a working prototype. Today's product is software: a phone, a stand, and the diagnostic engine.

What accumulates

Every paper makes the next correction sharper

A language model starts from zero on every request. A context window is not a memory: it is resent each time, it costs each time, and it is sealed inside one conversation. What we keep is a row per student that survives between calls — and it is the only part of the product that a better model cannot hand you.

What the system knows about this student

  • The chain rule error has now appeared three times.
  • Always on composite trigonometric functions, never on polynomials.
  • Solved unaided when the drill came the same evening; failed when it came three days later.

The model of the student deepens; the cost of a correction does not.

What it does differently

Stop drilling the symptom. Re-teach the decomposition, and schedule the drill the same evening.

What the tutor gets

The tutor no longer re-explains the chain rule from scratch. They are told this is a narrow, recurring gap on one function family — and that timing matters more than repetition for this student.