Pilot Case Study

Hindi Tutor

Will diaspora children practise Hindi at home,
on their own, with an AI?
A three-and-a-half-week pilot with a Primary 3 cohort at a leading Singapore Hindi school
27 March – 21 April 2026
1,249
Hindi sentences spoken at home
70%
Week 1 → Week 3 retention, no reminders
28
children, voluntarily, opt-in only
20
distinct topics explored

The host schoolA serious, curriculum-led institution

The pilot ran at one of only two organisations approved to teach Hindi as a Mother Tongue Language within Singapore's Ministry of Education framework — a long-established weekend Hindi school with over 3,000 enrolled students across its centres. The cohort for this pilot was a single Primary 3 group (children aged 8–9).

The cohort is a clean test of the heritage-learner profile: children who hear Hindi at home, are taught it formally once a week, and live the rest of their lives in a non-Hindi-dominant environment — in this case, English and Mandarin.

What Hindi Tutor aims to achieveThe gap is never the Saturday class. It is the other six days.

When it comes to Hindi for diaspora children the gap is rarely teaching — good weekend schools teach well. The gap is practice: a child gets two hours of Hindi on a Saturday and then lives the other six days almost entirely in English, and by the next Saturday much of it has decayed. Hindi Tutor exists to fill those six days — to be the daily speaking-practice layer that keeps a child's Hindi alive between classes.

Why Primary 3How to read the Hindi in this report

Primary 3 (ages 8–9) is the hinge year. The children are old enough to hold a real conversation and form complete sentences — which is what makes their speech measurable at all — and young enough that the slide into English-only replies has only just begun, so a daily habit can still take root. They can also drive a voice interface on their own, without a parent steering every turn.

One caveat

These are P3 children inside one of Singapore's MOE-recognised Hindi Mother-Tongue programmes — a relatively strong, formally-taught foundation. The Hindi on display here is what well-taught heritage 8–9-year-olds sound like. The more telling signal for a school is what happens at the other end of the same cohort: the children who arrived weak, or with no one at home to practise with — covered further down.

The question the pilot set out to answerThe hard problem is not pedagogy — it is practice volume

No teaching method or curriculum fixes the six days between Saturdays. So the pilot tested one narrow, honest question:

Will children practise Hindi at home, voluntarily, with an AI — without anyone pushing them to?

To make the answer meaningful, it was run under near-worst-case conditions on purpose:

What we deliberately did not do
  • No homework integration
  • No peer or social framing
  • No push notifications from us
  • Opt-in registration only — every parent had to sign their child up

The only contact was two emails from the school inviting families to register. Whatever engagement followed had to come from the child wanting to return.

What happenedAlmost all of the drop-off was at the front door

97 families in the cohort
97
Registered  (43%)
42
Began a conversation
28
Had 3+ conversations
25
Had 10+ conversations
10

The signal is the shape: the big loss is at registration, before the product is ever experienced. Of the 28 children who began a conversation, 25 came back for three or more. The leak is at the door, not inside the room.

The 55 we never reached

Families who never registered. The product never had a chance to reach them — the single biggest lever to pull.

89% stuck after starting

Of the 28 who began a conversation, 25 returned for 3+. The leak is at the door, not inside the room.

Children who started, kept coming back

This is the clearest result of the pilot. Of the children active in Week 1, 70% were still active in Week 3 (7 of 10), and 60% in Week 2 — with no reminders, no streaks, no nudges of any kind.

Week 1
10
children active
Week 2
6
returned · 60%
Week 3
7
returned · 70%
Benchmark

Education apps typically lose 85–90% of their starters by Week 3; even Duolingo, the category leader, sits around 28–40%.

Two things rule out the easy explanations:

It replicated
60% → 67%

A second signup cohort of roughly 15 children, joining in mid-April, produced the same early-retention shape (W1→W2).

It survived the trial ending

8 children kept using Kiki on their own after the pilot formally closed on 21 April — 30+ further conversations the next week, at a longer average session length than during the trial itself. Voluntary engagement after novelty has stopped being the explanation.

The children spoke a lot of real Hindi

229
conversations · ~100–110 min of pure child speech
96.5%
of tokens in Hindi · 92% of turns zero English
9.2
words per turn · 200+ turns past 15 words

Topics ranged from "my day" and "my family" to Diwali, Holi, Raksha Bandhan, festival foods, and talking to grandparents — half the returning cohort tried eight or more.

A map of where Primary 3 Hindi actually breaks

At the end of every conversation, Kiki flags the child's specific errors and walks them through the corrections. Across the pilot this produced 997 individual corrections — in aggregate, a real-time picture of where these children struggle, of a kind that is hard to get from classroom observation alone.

Grammar — verb forms, tense, particles50%
Hinglish vocabulary — English word for a Hindi one28%
Gender agreement — masculine/feminine mismatch22%
Sentence structure / incomplete<1%

The near-absence of sentence-structure flags is the curriculum caveat made visible: these well-taught children already form complete, well-shaped Hindi sentences. What they need is polish — verb endings, गा/गी/गे agreement, and the Hindi word for the thing they only know in English. Gender agreement, the quietest and most consistent gap at 22%, is the classic heritage-speaker error pattern — exactly what daily at-home practice is built to chip away at.

Early individual signal

Within three weeks, on the same topics, individual children were already producing cleaner Hindi: one child's average sentence length rose from 4.7 to 12.4 words per turn between Week 1 and Week 3; another's grammar errors per conversation fell by roughly 80%. These are early indicators in the children who returned, not a claim of measured fluency gain.

The full cohortWhen a child can't make the sentence, Kiki builds it with them

The same cohort included children who barely spoke Hindi, who fell back to English mid-sentence, or who had no one at home to practise with. For those children the design question is not "does Kiki grade them?" — it is "does Kiki keep them in the conversation and hand them the Hindi they are missing?" Three real exchanges from the pilot (names changed):

4-year-old answering in English
Kikiअच्छा क्यों लग रहा है? क्या हुआ? 🤔
ChildI I am happy today. Khush
Kikiतुम आज खुश हो? क्या बात है! 😊
Kiki never says "wrong." She folds the full correct sentence — मैं आज खुश हूँ — back into her reply and keeps going.
8-year-old collapsing into English mid-story
Childना, नहीं। मैंने बिल्ली के पास आई।
Kikiअच्छा! तो तुम बिल्ली के पास गई थीं? 🐈
Childमैंने इट स्ट्राइक टू बाइट मी।
Kikiओह! उसने तुम्हें काटने की कोशिश की? 🐈
Her grammar is wrong, then collapses into English. Kiki rebuilds each turn into correct Hindi while staying curious about the story — so the correction never lands as a correction.
9-year-old reaching for an English word
Childकुछ जोक्स
Kikiकुछ चुटकुले? 🤩 सबसे मज़ेदार वाला कौन सा है?
"Jokes" becomes चुटकुले, planted in context. Every such turn activates one more Hindi word in a child who had it only in English.

This is the behaviour that decides whether a heritage-learning product works for the whole class rather than only the children who were already doing fine — and it is exactly what a general assistant, built to correct or simply to answer, does not do.

In their own wordsReal sentences, spoken at home, choosing Hindi over English

गरम-गरम जलेबी अपने माता-पिता के साथ खाने में सबसे ज्यादा मजा आता है।
"Eating hot jalebi with my parents is the most fun."
Rohan, 8 · festival foods
दिवाली में मेरा पापा एक्सीडेंटली एक रॉकेट घर में ले आया!
"At Diwali, my dad accidentally brought a rocket into the house!"
Sanchar, 8 · Indian birthdays
जब मैं मेरा दादा-दादी के साथ होता हूं, तब वह सिंड्रेला की कहानियां सुनाते हैं।
"When I'm with my grandparents, they tell me Cinderella stories."
Manav, 9 · talking to grandparents

None of this is rehearsed dialogue. It is what daily practice looks like when a child chooses it.

What this pilot provesThree weeks answers the practice question. Fluency follows from the practice.

What three weeks can show
70%

of Week-1 children returned in Week 3 — no reminders, no nudges.

89%

of children who started came back for 3+ conversations.

8

kept using Kiki voluntarily after the trial formally closed.

Sentence length +164%, grammar errors −80% in three weeks, for the children who returned.

What three weeks cannot show

Fluency or measured accuracy growth. Heritage-language research (Polinsky 2018; Montrul 2016) consistently finds that grammatical-accuracy shifts in children who already speak a language conversationally require 3–6 months of sustained exposure, not three weeks.

That is not a weakness in the product; it is how language acquisition works for heritage learners. The practice habit — which is what was in doubt — is now demonstrated, and the accuracy gains follow from sustained use.

The unsolved problem in heritage Hindi has never been the class. It is the six days between classes — and for one cohort of children, for three and a half weeks, those six days were filled with Hindi the children chose to speak.
All figures derived from platform data for the pilot cohort, 27 March – 21 April 2026. Child names changed for privacy.
Benchmarks: Business of Apps / OneSignal, Mobile App Benchmarks 2024 (education category). Retention reported as conditional W1-active → W3-active.
Prepared by Shubham Gupta, Founder · Hindi Tutor · hindispeakingtutor.in