About us

We are building an AI English teacher for every government-exam aspirant in India.

English Exam Pro is an AI-powered exam-preparation company. Our AI teachers explain every mistake, talk a student through a topic by voice, and remember where each student is weak — for the English section of SSC, bank, defence, teaching and state exams.

One-to-one teaching works. It has never been affordable.

Every year, millions of young Indians sit government exams in which English is a scoring section — and a large share of them studied in Hindi or another regional language. What moves their marks is what a good teacher does sitting beside them: notice the rule they keep getting wrong, explain it in the language they think in, and make them try again until it sticks.

That kind of teacher is in a few cities, costs more than most families can spend, and goes home at night. So most aspirants prepare alone, with downloaded PDFs and question banks that can tell them they were wrong — but never why, and never what to do tomorrow.

We think AI can be that teacher, for everyone, at any hour — if it is built to teach rather than to answer, and if every fact it stands on has been checked by a person.

AI is the product, not a feature on top of it

Every part of a student's day in English Exam Pro — what they practise, how a mistake is explained, what comes back for revision, the questions themselves — runs through models we have built our product around.

The AI teacher

A tutor on every screen of the app. It reads the question in front of the student and their own answer, then teaches: the rule as a pattern, worked examples, the trap the examiner set, one line to remember — and a question for the student to try. It discusses a question's answer only after the student has attempted it.

Live voice teachers

Students call a teacher and talk a topic through out loud, with the working written on a shared board as the teacher speaks. Students choose who they learn from — Ved or Veda, each with their own voice and manner. Calls run on Google's Gemini Live, with a speech-to-text → language model → speech chain as a fallback that takes over mid-call without dropping the student. Being rolled out to students now.

A teacher that remembers

After each talk, a model reads the whole transcript and updates a short notebook on the student — how they like to be taught, what they were taught, where they slipped. A note is kept only if it quotes the student's own answers from that transcript, so the memory is evidence, not guesswork. The next session starts from it.

Mistake intelligence

Every wrong answer a student gives — in practice, in mocks, in the placement test — stays open in their mistake book until they get it right without guessing. After a set, a model reads the open mistakes and names the rule behind them in two sentences; the numbers in that read are computed by our server, never by the model.

Adaptive from the first ten minutes

The placement test is adaptive: each question adjusts to the ability estimate so far, so ten minutes gives a reliable picture. The daily plan is rebuilt every day from the student's own answers — weakest format first, old mistakes coming back before they are forgotten.

An AI content engine, with humans on the gate

Models extract questions from past papers, audit them blind against the answer key, write explanations, draft new questions on proven patterns, plan and draft lessons, write memory hooks for vocabulary and storyboard teaching videos. Nothing reaches a student until a subject expert has approved it — every question moves Draft → In review → Approved → Published.

Model-agnostic, with Gemini at the voice

  • Voice: live calls on Google Gemini Live over LiveKit, with an Indian-language speech chain as a fallback. The student's audio never touches our servers; our API issues the room pass, owns the teacher's prompt and tools, and meters every second.
  • Text: a provider-agnostic AI service with a fallback chain across Gemini, Claude and GPT models, so a feature keeps working when one provider does not — and each feature runs on the model that teaches it best for what it costs.
  • One brain, two voices: the spoken teacher and the written one share the same tools and student data, so what a student is told on a call and in the chat cannot drift apart.
  • Platforms: Android and the web, from one codebase.

Four rules for AI that students can trust

AI teaches; experts verify

Models write and explain at a scale no team could. A wrong answer key teaches a wrong rule for a month, so every question is approved by a person before it is published.

Never a number we cannot explain

No accuracy rate under ten answers — the app shows a dash. Projected marks, not a vague readiness score. A student should be able to check every number we show them.

Teach in the language the student thinks in

English or Hinglish, by the student's choice. For a Hindi-medium aspirant, understanding a rule in Hindi first is not a shortcut — it is how understanding happens.

Cost-aware by design

Every AI feature is metered, and every voice call records its tokens and estimated cost. That discipline is what lets one-to-one teaching reach students who could never pay a private tutor.

From one section to one teacher per student

We started with the English section because it is the most teachable section on the paper — the same rules and words decide it in every exam — and because it is where a Hindi-medium student is most often left on their own. Today we cover the English section of SSC, bank, defence, teaching and state exams.

Next, we are putting the voice teacher in every student's hands, teaching in more Indian languages, and turning each student's mistake book into practice written for them alone.