For English-medium schools · Class II–V

Stop telling parents their child “needs to work harder.”

Vaaani finds the exact sound, spelling, or grammar link each child is missing — and gives your teacher the evidence to fix it. Runs on the computers you already have. No names, no accounts, and nothing a child does is ever sent to an outside AI.

  • No LLM in the teaching decision — nothing here is generated by ChatGPT or any chatbot
  • Built on the mother tongue — Bengali has no /z/, so “zero” comes out “jero”. The engine knows.
vaaani · learner #217 · live decision
today’s focus

She knows “phone”, she knows the /f/ sound — time to join them

Two things she already knows, but hasn’t joined up yet. Fixing this one link also makes photo, graph, laugh easier.

Just the right challenge — about 68% likely to get it, with room to learn.

what she knows so far
the /f/ sound — solid0.81
spelling “ph” — still shaky0.34
the word “phone” — getting there0.52
how it comes out · still on paper — 7.2s, two tries
why it slipped last time
sound and spelling not joined up yet.62
hasn’t met it often enough.23
a habit from her home language.11
“The link between the sound and how it’s written isn’t solid yet — we’ll practise that exact pair.”

Every line the engine shows is computed, not generated. The numbers are its own estimates, drawn from what the child actually did — no ChatGPT, no large language model, anywhere in the teaching decision.

What lands on your desk

One line per child — including the ones nobody caught.

Not a score out of ten. For each child: the exact link that is missing, why it slipped past a good class teacher, and how the engine saw it anyway. This is the sheet your teacher works from on Monday morning.

Class V · ICSE · Bengali home language · 4-week cycle
The report format, filled in so you can read it. A pilot fills it with your own children.
ChildWhat is missingWhy the teacher never saw itHow Vaaani sees itWhat we do about it
Shuili Spelling and pronunciation have come apartsound–spelling link not joined up Corrected word by word, so it never showed up as a pattern phonogram map — one sound, five spellings Irregular-word set: photo · graph · laugh
Soumya Past-tense endings “-ed”rule known, not yet automatic The ending is subtle and the sentence still reads right morphology — walked · played · wanted are three endings, one spelling Contrast drills on the three endings
Siddhartha “th” lands as a dental stop — “three” → “tree”home language · Bengali has no /θ/ Inconsistent, and the meaning survives L1 contrast edge — predicted before it appears Minimal pairs: three/tree · path/pat
Md. Akram Knows the rule, cannot reach it in timecorrect on paper — 7.2s, two tries Read as hesitation, or as nerves answer latency and retries, not just right or wrong Short timed sets until it is automatic
Anirban /z/ comes out as /dʒ/ — “zero” → “jero”home language · Bengali has no /z/ Accepted as accent, so nobody teaches it L1 contrast edge · zoo · busy · rose all inherit it The /z/ set, then the words it unlocks
Saroj Errors appear only when the task gets heavierholds up alone, breaks under load Looked random, so it was put down to carelessness self-repair and retry pattern across a session Rebuild the weak link before adding load
Sinjini Plural “-es” after /s, z, ʃ/has not met it often enough Reading is fluent, so the gap stays hidden rule-level evidence, not word-level boxes · buses · dishes as one rule
0 LLM calls

In any teaching decision. Every line above is computed from what the child did — no ChatGPT, no chatbot, no child’s work sent to an outside AI.

4 studies on the engine

We test our own reasoning before we ask a school to trust it: does tracing the links between sounds, spellings and words actually predict what a learner gets right next? Run on the shipped engine, repeatable, with the results as they came out. Read the method →

Where the evidence honestly stands: the engine, its reasoning and its self-checks are built and reproducible — but Vaaani has not yet run a full classroom pilot. There is no school logo to show you and we are not going to invent one. Your school can be the first, on 30 students, free — and we will publish what the numbers say whichever way they fall.

Why schools choose us

You can defend a decision you can see.

Most classroom AI is a large language model in a wrapper — helpful, but you cannot say why it taught what it taught. Vaaani was built the other way round.

Typical AI tutor

A black box

  • A chatbot decides what to show — you can’t see why
  • “Personalised”, but nothing you can actually look at
  • Hands over the answer; the thinking stays hidden
  • Never checks whether it was right
  • Sends the child’s work off to a cloud
Vaaani

A glass box

  • Every lesson is chosen for a reason it shows you
  • Keeps a living picture of your child — open it and read it
  • Tells you why it chose, and how likely it is to work
  • Checks its own guesses against what really happened
  • No name, no account — never sent to an outside AI

Inside the glass box

Five instruments, one growing child.

No large language model anywhere in the teaching decision. Just an inspectable, estimated state and explicit reasoning — the kind a linguist, not a chatbot, would use.

The Cognitive Twin

A living picture of what your child knows — it updates with every answer, and fades gently over time, the way real memory does. Every lesson is picked for your child alone, so no two children follow the same path.

‘tri-’ words (three)0.74
‘aqua-’ words (water)0.41

CASCADE — a web, not a list

Words are learned in a web, not one at a time. Master one link and the words next to it get easier — learn the /f/ in “phone” and photo, graph, laugh come along for the ride.

phonephotographlaughtough

The Cause-net

When your child gets something wrong, it works out why — the spelling isn’t linked to the sound yet, not enough practice, or a habit carried over from the home language — then fixes that exact thing.

home-language habit.58
not enough practice.27

Calibration — it keeps score

Before each lesson it predicts how your child will do, then checks itself against what actually happened. A teacher honest enough to grade its own guesses.

guessed 70% · she scored 68% · honest

Fluency — paper vs. real time

Knowing a rule and using it in the moment are not the same thing. This watches how the answer arrives — fast and sure, or slow and second-guessed — and tells effortful recall apart from real command. A correct answer that took eight seconds and two tries isn’t mastered yet, and it says so, so a child is never marked “done” on a skill they can only manage on paper.

in real time — fast, first try0.92
on paper — slow, self-corrected0.34
answered in 1.8s · first try · automatic — really hers now

The unfair advantage

Bridged from the mother tongue.

An Indian child learning English is not a blank slate — they already own an isomorphic web in Hindi, Bangla, Tamil. Vaaani grafts that head start onto the English graph, and reads L1 into why a sound is hard.

A Bengali speaker saying “jero” for “zero” isn’t making a mistake — Bengali has no /z/. The engine knows that, and teaches from it.

“Why do Bengali speakers often say jero instead of zero?”
Bengali doesn’t have the /z/ sound, so the brain maps it to the nearest one it owns, /dʒ/. That’s your mother tongue’s phonology at work — not an error. Now: which English sounds does your first language lack? Let’s find them.

Run a pilot

See it work on your own students.

We start small and honest — a single cohort, real children, measured results you can read line by line.

01

Pick 30 students

One class or centre, free. Set the home languages. No installation — it runs in the browser, on a phone or a lab machine.

02

Children do lessons

Each child works through short lessons. The picture of what they know fills in; every choice records why it was made and how likely it was to land.

03

Read the evidence

You see what each child knows, why they slipped, and how well the engine’s own guesses held up — not a mystery score.

Where we honestly stand: no classroom pilot has run yet — the first one is the offer on this page. The engine’s starting numbers come from published research and get sharper as real pupils use it. We show what we’ve actually measured, and flag anything that is still an estimate.

Bring a tutor that can explain itself to your school.

Thirty students, free, and a diagnostic sheet you can hand to a class teacher. We set it up personally.