India's own exam-AI model
India runs on entrance exams. We built the model underneath.
DhiX AI-1 is our own vision-language model for Indian exams: it reads a typed doubt or a photograph of one — the diagram, the printed problem, the four lines a student already tried — and teaches from what is on the page. It learns from a corpus we built ourselves with expert educators, and it reaches students two ways: our own consumer brand, and the institutes who put their name on it.
- Modality
- Vision-language
- text and images in one model — a typed doubt, a photographed question, a diagram, a page of handwritten working
- Trained on
- Our own exam corpus
- Indian exam content curated with expert educators and verified by hand
- Reasons at
- JEE and NEET level
- works a multi-step problem the way an Indian teacher does, and marks it the way the paper does
- Improves from
- Every verified correction
- each correction our educators make goes back into the model
The market
One exam decides a life. Millions sit one every year.
In India a seat at a top institution runs through a single high-stakes paper, and the whole household organises around it. That is not a segment of an education market — it is the market, and it repeats every twelve months with a new cohort.
These populations overlap — one student sits boards and more than one entrance exam in the same year — so they describe reach, not a summable market. Figures current as of August 2026; every one is sourced.
The spend behind it
India's edtech market, 2024 to 2030[5]. Test-prep is a multi-billion-dollar segment of it, compounding at roughly 9% a year[6] and moving online fast.
The money is already being spent. It is being spent on the wrong thing.
Families pay for coaching hours, printed modules and recorded lectures — inventory built for a classroom of two hundred, sold to one student at a time. What none of it does is answer the question in front of the student at eleven at night, in their language, and be right about this year's paper. That is a software problem, and until recently nobody could solve it.
Why now
The capability arrived. The trust did not.
The category leader collapsed: BYJU'S fell from a ~US$22B peak in 2022 to effectively zero by 2024. What is scarce in Indian exam prep today is not content and not distribution — it is a reason to believe an answer. Three things changed at once, and they are why this is a window rather than a plan.
CNBC · TechCrunch, 2024
Value moved from the recording to the model
A library of recorded lectures was the asset of the last decade — expensive to make, and identical for every student who watched it. What a student will pay for now is something that answers their question, about their paper, at the moment they are stuck. The inventory that used to be the moat is becoming the commodity.
India is not a translation problem
Its exams have their own syllabi, their own eleven question formats, their own marking schemes, their own cut-offs and their own teaching languages. A model built for everything was never taught any of it, and cannot be prompted into knowing it. The specificity that makes this market hard to serve is exactly what makes it defensible once served.
Owning a model stopped being a moonshot
Training a capable model on content you own is now a company-sized undertaking rather than a research-lab one. That is what turns an Indian exam model from an ambition into a plan — and it is a window, because the same fall in cost is available to whoever moves next.
What we own
Four assets. None of them rentable.
Cheaper, better general AI helps us — we would rather that whole field kept improving, and we would rather not be in the business of matching it. What does not commoditise is underneath: a model trained on Indian exam content, the content itself, the product that delivers both, and two channels to a student.
DhiX AI-1
Our own vision-language model for Indian exams. It reads what a student is actually looking at — a photographed problem, the figure beside it, four lines of their own handwriting — reasons through a JEE- or NEET-level question the way a good teacher does, and marks it the way the paper marks it. Every correction our educators make goes back into it. It is a model we trained, on a corpus we built, that we can run where we choose and improve on our own schedule.
The content it learned from, and students learn from
The same library does two jobs. 160,000+ expert-authored questions across 465+ chapters — worked solutions, the real figures from the papers, the marking rules of eleven Indian question formats — power practice and mocks on the live product, and every reviewed solution is training material for the model. An open-web scrape of coaching PDFs is not this, and a model trained on one behaves exactly like the assistants students already distrust.
A whole product, not a demo of a model
Doubt-solving, an AI voice tutor a student can call, timed mocks rendered as the real computer-based test, rank and college prediction from official data, weekly parent reports, and the console an institute runs its cohorts from. All of it live. A model without this is a research result; a product without the model is a wrapper.
Two ways to the same student
Our own consumer brand sells directly to aspirants, and institutes run the identical platform under their own name and their own database. The first external client is live. The second channel puts our content in front of students we would never have acquired ourselves — and both channels feed the same corpus.
The product
What the model does all day.
Six surfaces, all live today. The model is the asset; this is the shape it reaches a student in — and the reason an institute can put its name on the whole thing rather than on a feature.
Doubts, from a photo
Snap the problem, the figure, or your own half-finished working. Checked before it is shown.
An AI voice tutor
Named AI subject tutors, speaking Hinglish, interruptible mid-sentence, writing on a live board.
Mocks on the real screen
The computer-based test as it is actually sat, with the paper's own marking scheme.
Rank and colleges
An All-India Rank with a band around it, from official cut-off and counselling data.
A weekly parent report
Computed from real attempts. No number in it is written by a model.
An institute console
Cohorts, staff, content and analytics for the client whose brand is on the product.
Defensibility
It gets harder to catch the longer it runs.
A model that is licensed can be licensed by anyone, and a product built on one is a quarter of engineering away from being copied. Ours improves from something a competitor cannot start with: students using it, and educators correcting it.
Students and institutes use it
two channels, one product, every doubt and every attempt on the same content
Our educators correct what is wrong
a wrong key, a weak explanation, a figure that does not match the paper
The corpus and the model absorb it
the fix lands once, and DhiX AI-1 learns from it
Answers get better than a general model's
on this syllabus, this year's pattern, this marking scheme
And round again. Better answers bring more students and more institutes, which is what produces the next round of corrections. Every turn costs us less than the one before and leaves a competitor a year further behind.
The parts nobody can buy
Exam content written by people who teach these papers
Not a licensing deal and not a scrape — an editorial operation, run continuously, by educators who know how an Indian paper is actually marked. Money alone does not shorten it; you also have to run it for years.
An official rank archive that only ever grows
Cut-off curves per exam, category and year, and five years of official counselling closing ranks. It is public data that nobody else has bothered to gather, clean and keep current — and a sixth year arrives whether or not a competitor started.
The record of what students actually get wrong
Only real attempts reveal how hard a question truly is and which misconception sits behind each wrong option. That accrues to whoever is already serving students, and to nobody else.
Stated once, and bounded: as of August 2026 we are not aware of an Indian exam-prep company serving students from its own exam model, trained on a corpus it authored itself. We would rather be corrected on that than assert a negative about a whole market.
The business model
One asset, sold twice.
The same model, corpus and product monetise through two independent channels — direct subscriptions from students, and platform licences from institutes who put their own name on it. The first external client is already live on the second.
D2CStudents
Students subscribe directly to our own brand.
- Exam intent is the acquisition channel — the rank and college predictors bring a student in, the tutor is what keeps them.
- Recurring subscription revenue against a purchase decision the household has already made and is already funding.
- Every session improves the asset: the content gets corrected and the corpus behind the model grows.
B2BInstitutes
Institutes run the identical platform under their own name.
- Platform licence revenue, contracted per institute, against a budget that is currently spent on people and print.
- Their brand, their domain, their own database — and a licence scoped to the exams they actually teach.
- It reaches students we would never have acquired ourselves, on content that is already built.
One product, every deployment
There is no per-client fork to maintain, so a fix reaches everyone the day it ships. Branching the product per customer is the usual reason white-label quietly kills a roadmap.
The asset is shared; the students are not
One content library and one model serve every deployment, while each client's students and results stay entirely their own. A second client adds marginal cost, not a second content operation.
Each line makes the other cheaper
The consumer product proves what works before an institute ever sees it, and institute cohorts widen the corpus that improves the consumer product. Neither line pays for a separate engine.
Clients are charged from a published rate card that is the same for every customer, and what an answer costs us to produce stays on our side of the wall — so improving our own economics never appears as a surprise on somebody's invoice.
Shared in the data room
Named here, numbered privately. Revenue, retention and pipeline are reviewed under NDA with the deck — we do not post them, and we do not estimate them on a public page.
Where this goes
Built as an exam engine, not an exam list.
Seven exams are what we serve today; they are not what we built. No part of the platform is written around JEE or NEET and the model learns from whatever corpus we put in front of it, so the ceiling here is the number of exams India runs — and then the number of things India studies.
Seven exams, live today
- JEE Main and JEE Advanced
- NEET-UG
- CBSE boards
- Foundation (Class VIII–X)
- Banking (IBPS)
- SSC
Serving real students today, on both channels.
The next exams
- State engineering and medical CETs
- CUET
- UPSC, and the deeper SSC tiers
- Further government and banking cadres
Roadmap, not shipped. Each is content plus a catalogue entry — the same work, repeated.
Beyond exams
- School tutoring across the regular academic year
- Skilling and professional certification
- Vernacular teaching, in the language a student thinks in
Roadmap, not shipped, and deliberately further out. Same engine, a different body of content.
Said plainly, because roadmaps are cheap: only the first column exists. The other two are the direction the architecture already points in, and we would rather show you the direction than pretend seven exams are the whole ambition.
Status
Early on traffic. Not early on product.
We built the engine first, deliberately, and we would rather show you the asset than dress up a user number you will check anyway. Here is what is running today.
Two products, one build
Our own consumer brand, and the first external white-label client on its own domain and its own database. A fix we ship reaches both the same day.
The whole product
Doubt-solving, an AI voice tutor, timed mocks on real exam screens, rank and college prediction, parent reports and the institute console.
Our own model
DhiX AI-1 reads typed and photographed questions, reasons over JEE- and NEET-level problems, and improves from every correction our educators make.
This page carries only what is independently verifiable. Financials, cohort data and the engineering handbook behind every claim are shared under NDA.
The vision
India's own model for the exams that decide Indian futures.
Every serious exam market ends up with an intelligence built for it rather than borrowed. We intend that to be ours: the model, the corpus underneath it, and the product that puts both in a student's hands — starting with seven exams, and not stopping at exams.
The ask
Raising to widen the corpus and take the model to scale.
A priced round on a product that is already live: our own exam model, an expert-authored bank behind it, official rank data, and a first external client serving real students.
Deck and terms on request. Commercial metrics, cohort data and the engineering handbook behind every claim on this page are reviewed privately, under NDA.
Compute and the people around it, so DhiX AI-1 covers more subjects, more exams and more of what a student can put in front of a camera.
More senior educators, more subjects, and a named human verdict on every question in the bank rather than on most of them.
Sales, onboarding and success to turn a shipped white-label platform into signed licences — and the capacity to carry them.
Or write to us directly: founders@dhixai.com
Sources
Figures current as of August 2026.
- 1.NEET (UG) 2026 candidates. NTA press release, 16 July 2026 ("NTA Declares Result of NEET (UG) 2026"): "close to 20 lakh candidates appeared", "11.21 lakh candidates have qualified".
- 2.JEE Main / JEE Advanced 2025 registrations. NTA cycle figures for JEE Main, and the record ~190K JEE Advanced registrations reported for 2025 (ALLEN).
- 3.CBSE Class XII 2025. CBSE's own 2025 board-exam cycle figures.
- 4.SSC CGL 2024 applications. SSC's 2024 CGL cycle, reported via CareerPower. The wider govt-exam figure (SSC + banking + railways + UPSC) is directional.
- 5.Edtech market size. India edtech ~US$7.5B (2024) → ~US$29B (2030): IAMAI–Grant Thornton Bharat report, via IBEF, 2024.
- 6.Test-prep growth. Multi-billion-dollar segment at ~9%+ CAGR: Technavio, 2024 (vendor estimate, directional).
- 7.Incumbent landscape. BYJU'S peak-to-collapse (~US$22B → ~0): CNBC and TechCrunch, 2024.
Figures above are public, third-party estimates cited for context. Market sizings are directional vendor forecasts, not audited data. DhiX AI's own commercial metrics are shared, verified, under NDA with the deck.
Our own figures. The question-bank count is what a student can actually be served, measured 20 August 2026 and rounded down. Chapter coverage and the counselling archive come from the same measurement. Each is reproducible from the repository, and the method, sample size and dates travel with the deck.