lipi

Lipi — Devpost copy (paste as-is) Tagline: Ten minutes, in their own language, before anyone decides what they are.

Track: Ideathon

Inspiration My cousin failed Telugu twice.

Not English — Telugu. The language he speaks at home, the one he argues with his mother in. He could hear a lesson once and retell the whole thing. He just could not get it off the page.

The school decided he was careless and gave him extra homework, which is roughly the worst thing you can hand a child who cannot decode text in the first place. By class 6 he had decided he was stupid. That belief cost him far more than the reading difficulty ever did.

Nobody in that building was cruel. They had forty-five children in a room and no way to tell the difference between a child who will not read and a child who cannot. That gap is a screening problem before it is anything else, and it is the one I want to close.

What it does Lipi is a ten-minute game a six-year-old plays on a cheap classroom tablet, in Telugu or Hindi or Tamil, that ends with the teacher knowing which three children in that room need a closer look.

Five tasks, each a game to the child and a validated construct underneath: rapid naming, sound isolation, akshara-sound matching, word chains, and one short read-aloud. Everything is timed to the millisecond.

Three outputs:

The teacher gets a class list ranked by risk, and one plain sentence per flagged child that ends in an action — "reads accurately but very slowly, give extra time before you give extra homework." The parent gets one page in their own language, no jargon, opening with what their child is good at, and a phone number for the nearest assessment centre. The district gets anonymised prevalence data — exactly the number every state FLN mission is supposed to report and currently cannot. Lipi outputs green / amber / red. It never prints the word "dyslexia" on a child's report. Diagnosis is a clinical act that needs a qualified psychologist, and any product pretending otherwise deserves to be shut down. Lipi flags risk and hands it to a human.

How we built it This is an Ideathon entry, so what exists today is a fully worked plan and a designed interface — not a shipped app. I would rather say that plainly than imply a build that isn't there.

What I did build: the full task battery design, the scoring approach, the architecture, four screen mockups, the business model, a costed rollout timeline, and a risk register. All of it is in the project plan PDF.

The intended stack, and why:

Offline-first, Flutter on a Rs 6,000 Android tablet. The schools that need this most have the worst connectivity. If Lipi needs a signal to score a child, it doesn't work in the places it exists for. The screening has to complete in aeroplane mode. Speech on-device (distilled Whisper-small). Worse than a server model, and I don't care — the alternative is shipping six-year-olds' voices over the internet. I need phoneme timings and pause structure, not a perfect transcript. A deliberately boring scorer. Gradient-boosted trees or regularised logistic regression over ~40 features. Not a deep net. When a teacher asks why a child was flagged I have to be able to answer, and a state education officer has to be able to audit it before it goes near a government school. Adaptive item selection (IRT) so a fluent child finishes in six minutes and a struggling child gets more items where the information is. Challenges we ran into The hard one was realising every existing screener is built for the wrong writing system.

English spelling is irregular — through, though, tough — so a child memorises chaos, and English dyslexia shows up loudly as accuracy errors. You catch it by counting mistakes.

But Telugu, Devanagari, Tamil and Kannada are largely phonetic. The symbol tells you the sound. So a dyslexic child in a transparent script often reads accurately but painfully slowly, sounding out every akshara, never becoming fluent. An accuracy-based screener looks at that child, finds almost no errors, and passes them.

We are not under-detecting because we lack tools. We are under-detecting because we imported the wrong ones. That single fact redesigned the whole product — it is why rapid naming is the load-bearing task and why every response is timed.

The second challenge was ethical and I couldn't engineer around it. Screening creates referrals a district may not absorb. My own district mockup shows 1,847 flagged children and three assessors — 26 months to clear. Hiding that would make Lipi part of the problem, so the product reports the bottleneck to the district and ships classroom accommodations alongside the flag, so an amber child gets extra time even if no assessor is free for a year.

The third is technical and unresolved. Child speech recognition is genuinely bad — high pitch, regional accents, code-mixing, ceiling fans. So I designed the battery so four of the five tasks need only tap input and timing, which the device measures perfectly. The product survives the failure of its hardest component. That was on purpose.

Accomplishments that we're proud of Finding the orthography insight. Research-backed, non-obvious, and it explains why the problem persists despite good tools existing elsewhere. A plan shaped like a company, not a demo. Real buyer, real price anchor (Rs 30 vs Rs 3,000+), and a moat that isn't code — the validated item bank and per-language norm tables. Ethics designed in, not bolted on. No diagnostic language anywhere. Reports lead with strengths. Audio never leaves the device. A human always decides. Being explicit about the asymmetry. A false negative is a child called lazy for another year; a false positive is one unnecessary assessment. Those costs aren't equal, so the threshold is tuned for sensitivity on purpose — and a district can override it. Saying what I haven't got. No child has touched this yet. Every projection is marked as one. What we learned That the interesting question was not "can dyslexia be predicted" but "can it be caught early enough that someone can still act." Those produce completely different products.

That the moat here is not software. Anyone can write the app. The expensive, slow, unglamorous part is the validated item bank and norm tables built from real Indian children — a linguist, a special educator, and a study. Which is precisely why nobody has done it.

And that honest limits belong in the pitch, not the appendix. I can build the entire technical side; I have no clinical authority and no validation cohort, and no amount of engineering substitutes for either. That's why the first hire is a special educator, not a developer.

What's next for Lipi Months 0-3 — Build the Telugu battery with two special educators. Paper prototype, 30 children.

Months 3-6 — Working app in 3 schools, ~300 children. Does the game hold a six-year-old's attention?

Months 6-12 — The validation study, ~1,200 children, scored against real clinical assessment. This is the gate. Without it Lipi is a nice app; with it, it's an instrument a government can buy.

Year 2 — Hindi and Tamil, 50 paying schools, first district pilot.

Year 3 — State contract, six languages.

Then: screening into structured support, so the flag leads somewhere. Progress monitoring each term. Dysgraphia via handwriting photos and dyscalculia via number-sense tasks, on the same rails.

Eventually beyond India — the transparent-orthography problem is structural, not Indian. Spanish, Italian, Turkish, Indonesian and Swahili have the same gap and the same missing tool.

Built With

• accessibility
• built-with-concept
• education
• flutter
• item-response-theory
• offline-first
• on-device-ml
• scikit-learn
• whisper