VED AI: YOUR MENTOR,
ONE QUESTION AWAY

By 2025, asking an AI anything - at any hour, in whatever language you think in - had become normal for students. That part was solved, and not by us.

What it didn't solve is more interesting. A general assistant can answer any question and has nothing to hand you afterwards. It doesn't know your class, your board, your target exam or your last test score, and it has no video, no notes and no paper of its own to give you.

Vedantu had the exact opposite problem. All of that material already existed - across four products - and none of it could be asked a question.

An enormous content library, and years of data about every student - class, board, language, last week's score. All of it sitting behind a search box or a menu, waiting to be found by someone who already knows what they're looking for.

So the question wasn't "how do we compete with ChatGPT." It was simpler and more annoying: why can't a student just ask us?

Before the decisions, what the thing actually is. Ved is a chat assistant inside Vedantu's app. A student asks the way they'd ask a tutor - typed or spoken, in English, Hindi or Hinglish - and gets an answer written for their class and board, followed by the Vedantu material that goes with it.

It isn't a general assistant with a school theme painted on. Three things sit behind every single reply: who the student is, how Vedantu's syllabus is structured for that board and class, and Vedantu's own content and performance data about that student. Every decision in this section follows from those three being available - and from the fact that they're worth nothing unless a student can reach them in one sentence.

In its first year it carried roughly a million messages from about three hundred thousand students, across the Android and iOS apps, mobile web and desktop.

The third screen is the whole argument: identical words from a class 8 and a class 12 student, two different replies.

When a student asked for something bigger than one answer - "I want to study all of biology basics, any playlist?" - a block of text was the wrong object. Ved returned a card instead: Learn this topic with Ved · Video · Notes · Questions.

With an assistant, a large share of the experience isn't laid out - it's written. How long a reply runs, what tone it takes, when it asks a question back, and what it refuses to do are all design decisions, and they were specified as precisely as any layout.

A tutor who answers a twelve-year-old the way they answer a repeat-year engineering candidate is a bad tutor.

Underneath the tiers, four rules that held everywhere. Keep replies the length of a message, not an article. Give detail gradually rather than dumping a plan. Ask at most one question back, and only when the answer genuinely can't proceed without it. End with encouragement and one concrete next step.

An assistant built for teenagers under exam pressure meets things it must not attempt. Rather than leaving this to the model, each case got a defined response.

Underneath the tiers, four rules that held everywhere. Keep replies the length of a message, not an article. Give detail gradually rather than dumping a plan. Ask at most one question back, and only when the answer genuinely can't proceed without it. End with encouragement and one concrete next step.

Two things went wrong in ways no screen could fix. Ved sometimes answered a student who wrote in Hindi with a reply in English. And it sometimes answered a question about this year's exam with last year's information - dates, cutoffs and eligibility rules all move.

The PM and engineering team led both fixes. What they changed for me was what I could assume: I couldn't take it as given that a reply would come back in the language the student wrote in, and I couldn't treat anything dated as safe to display without a check behind it.

The obvious response to flat engagement is to build more. The data said the opposite. Ved was already doing four genuinely valuable things and students couldn't see any of them.

So the second version is mostly not new capability. The design file is organised the way the work actually ran - UI changes on one side, new features on the other, mobile and web tracked separately - and the UI-changes side is roughly three times the size of the new-features side. What got added was memory, a library, a history, and a way to ask without typing. What got changed was everything about how a student finds any of it.

The first version opened on an empty input box. That box is the one-message finding in a single object: it asks the student to already know what Ved can do before they can get anything out of it.

The library solved the content problem. It didn't solve the thread problem: a student who worked through something good with Ved on Tuesday had no way back to it on Friday.

Empty and broken states are where a product either keeps a student or loses them, and they're the first thing that gets skipped. Ved is used at midnight on patchy mobile data, by someone who has a test in the morning - so these four got the same attention as the happy path.

Each one does a job rather than apologising. The generation error offers the retry inline, without throwing away what's already on screen. The offline state says which side the problem is on, so the student doesn't blame their own answer. The empty library sends them somewhere to fill it. The empty bookmarks teach the gesture, because a feature nobody knows how to trigger is the same as a feature that doesn't exist - which is the lesson the whole rebuild came from.

In the first version, opening a chapter page took the student to a new screen. On a phone that's the only sensible option. On the web it isn't, and a full page change meant losing the conversation you were in the middle of.

The second version brought all content into a side panel on the web. Same reasoning both times, two different answers: never make a student lose their place to look at something.

Two findings pointed the same way. Use peaks late at night, on a phone. And a large share of students write in Hindi or a regional language, where typing means switching keyboards or spelling their language out in English letters - friction that speaking removes entirely.

Voice exists in two forms, and they aren't the same decision. The microphone in the chat box is a shortcut: speak instead of type, get an ordinary reply. Live mode is a continuous spoken conversation with its own states.

The largest single thing students asked for was help with tests and scores. In the first version that meant being sent somewhere - a test series, a practice paper, another product.

Now quiz me on the second law is a sentence Ved can act on. It builds the questions from the same syllabus data the rest of the product runs on, marks each answer as it's given, explains the ones that went wrong, and closes with a summary of what to revise. The whole thing happens inside the thread that raised the topic.

The last move was leadership's: give the four capabilities names, so a student can point at them instead of guessing them into existence. The naming wasn't mine. Making each one a real thing, in the place the student already is, was.

Because once a capability had an identity, it could leave the chat entirely. Ved isn't a product sitting next to Vedantu's other products. It runs inside them.

The same student profile, the same syllabus knowledge, the same memory - reached from wherever the student happens to be. A student in a live class doesn't open Ved. Ved is already in the class.

This is also the part a general assistant structurally cannot do. It can't sit inside a live class, read a test report, or reach a question bank it doesn't have.

WAVE is Vedantu's live and recorded class platform. One teacher, hundreds of students, and doubts arriving from all of them at once. The teacher answers one at a time. Everyone else waits, and a waiting student stops watching.

LectureLens sits under the player: notes for the class, a quiz, tests, jump points for every topic covered - and doubt-asking built into the session itself.

A score report tells a student what happened. It rarely tells them what to do. ScoreBooster lives inside the report itself, so the moment a student sees where the marks went, the way to fix it is in the same place - no separate product, no second search.

It reads the test: what was scored, which questions went wrong, which topics are weak, and what to work through next to move up.

Typing a physics problem with three subscripts and a diagram is worse than being stuck. So the student photographs the question and gets it worked through step by step, with the concept behind each step and the resources that go with it - the notes for that topic, the video, similar questions from past papers.

Doubt experts are still there, and connecting to one stays the headline path. InstaSolve doesn't replace that; it means a student isn't sitting on their hands while they wait, and it handles the cases where a full expert session was never needed.

The chapter path from the first version, now something a student can reach for by name instead of stumbling into. Watch the video, read the notes, test yourself - matched to your class, your board and your language, and short enough to finish a topic in one sitting.