Apurv Mehra | Portfolio

We started Disha AI with a simple premise: high-stakes exam aspirants in India don't lack content - they lack affordable, personalized guidance. If we could build an AI tutor that delivers structured feedback, contextual practice, and continuous nudges at a fraction of the cost of a coaching institute, we could unlock access for millions.

We chose UPSC as our first wedge.

Why UPSC?

UPSC is one of India's most aspirational exams. Around 4-5 million aspirants are preparing for it at any given time, and the coaching market around it runs into billions. But more importantly, the exam is overwhelmingly text-based — history, geography, polity, economics, current affairs — all domains where LLMs perform well.

Compare this to JEE, where questions involve geometry diagrams, organic chemistry mechanisms, or biology illustrations that require visual reasoning. Or competitive math, where spatial and diagrammatic explanations are essential. UPSC's syllabus — heavy on analysis, synthesis, and written articulation — felt like the ideal playground for an AI tutor given current performance of LLMs.

We had also learned from our earlier experiments in professional skilling that motivation was a persistent challenge. The core issue was that people are motivated by fear or clear outcomes — if there's a tangible reward like a job at the end, they'll push through. Otherwise, only the few who genuinely enjoy learning stick around; the rest drop off. Since we weren't in a position to promise jobs and were only offering certificates, we tried a different approach — making the experience more bite-sized and gamified. This did improve engagement outcomes, but not to the extent we had hoped for.

Our Four Hypotheses

Going into this, we had a clear set of bets:

Vertical depth wins. If we built features that solved UPSC-specific needs exceptionally well — news analysis, mains answer evaluation, personalized feedback on subjective writing — students would value it over generic AI tools.
AI-first preparation is viable. With rapid model improvements, we believed some students could rely entirely on an AI tutor for their preparation, without needing a traditional coaching institute.
High-stakes = high motivation. Unlike skilling or casual learning, test prep is inherently driven by urgency. The exam date is fixed, the syllabus is defined, and the stakes are life-changing. We expected this to solve the retention problems we'd seen in other verticals.
Scalability is cleaner in test prep. In our skilling work, we kept running into a hard question: how do you prove the outcome? Getting someone a job is a complex, multi-variable problem that's hard to attribute to any single platform. With test prep, the outcome is the exam itself. Students understand that preparation requires hard work, coaching tools, and practice — and the value chain is much clearer.

The Raghav Moment

We had spent weeks building a flashy new website for Disha AI. It had a polished landing page, an AI persona called "Disha" who was supposed to be a friendly, always-on tutor, and deep features for UPSC preparation. We went live, and our co-founder Kashish — who has close to 300,000 followers on Instagram and is himself an IAS officer turned entrepreneur — started putting out content.

The traffic was encouraging. We were getting 1,000+ daily hits. From Instagram, we funneled about 10,000 engaged users into Telegram groups, and from those, we had a "Super 30" cohort of ~35 highly motivated aspirants.

But here's the thing: fewer than 1% of visitors were actually onboarding onto the platform.

We didn't have contact details for most visitors, so we started reaching out to users from our Telegram and Super 30 groups. One conversation with a student — let's call him Raghav — was particularly revealing.

"Hey Raghav, hope your preparations are going well. Wanted to check — did you get a chance to try the Disha AI platform? We'd created a personalized question set based on your DAF."

"Yeah, I opened the link. It took me to some website with this AI Disha. I honestly thought it was Kashish sir's platform and I was expecting Kashish sir. But when I saw it was just another AI chatbot thing, I looked at the landing page and dropped off."

This conversation played out in different flavors with different students, but the essence was always the same. For a high-stakes exam like UPSC, aspirants want a relatable human face — someone they can trust. An AI tutor, no matter how well-built, doesn't bring that trust by default.

The students who did use the platform were the ones who specifically knew that Disha AI was built by Kashish Mittal — a person they already trusted. Without that human anchor, the platform felt like "yet another generic AI tool" in a world already saturated with ChatGPT and Gemini.

What We Got Right, What We Got Wrong

After 9 months of building in this space, here's how our hypotheses held up.

Vertical depth matters — but content is table stakes

We built solid UPSC-specific features. The mains answer evaluator gave detailed, structured feedback on subjective answers. The current affairs engine delivered daily news analysis mapped to the UPSC syllabus. Users who engaged with these features found genuine value.

But we underestimated the sheer volume of learning content required. UPSC preparation isn't just about practice and feedback — there's a massive "learning phase" where aspirants need to absorb foundational knowledge across dozens of subjects. That phase is still largely driven by human teachers, because students want to learn from a recognizable, trusted face. Content creation at that depth and quality turned out to be a much bigger lift than we'd anticipated.

AI alone can't drive the experience

This was perhaps our biggest learning. AI adds real, incremental value to educators: generating additional practice questions, providing instant feedback, offering better examples and explanations. But AI cannot be the primary driver of the learning experience — at least not today.

We experimented with AI-powered learning sessions, trying to get the tutor to guide a student through a 30-45 minute study session. It couldn't sustain engagement. It couldn't keep the discussion anchored to specific learning outcomes. And critically, it's sycophantic by nature — it struggles to provide the tough, honest push that a good teacher gives when a student is slacking or going off-track.

The future isn't AI replacing teachers. It's AI amplifying teachers.

Motivation is still the hardest problem

We had assumed that high-stakes exams would solve the motivation challenge. The reality was more nuanced.

When we looked closely at our ICP (ideal customer profile), only about 10-15% of UPSC aspirants are genuinely serious about their preparation. A large number lack the foundational knowledge — basic mathematics, history, geography from their K-12 years — that the exam demands, making the preparation feel overwhelming. Many aspirants gravitate toward UPSC not out of deep conviction but because they don't know what else to do next.

The behavioral pattern we saw was eerily similar to gym memberships. Sitting with books and actively studying for hours is hard. Turning on YouTube and passively watching lecture videos is easier. People gravitate toward the path of least resistance, and only the most disciplined actually maintain a consistent preparation rhythm.

Our retention funnel told the story starkly: out of ~900 students who started learning on the platform, only 9 were still on track after 30 days. A 99% drop-off.

High stakes cuts both ways

Here's the most counterintuitive lesson. We'd hypothesized that high-stakes exams would mean higher motivation. But high stakes also mean higher trust requirements.

When the outcome really matters — when your career and future are on the line — you don't want to experiment with an unproven AI tool. You want to go to the coaching institute with a track record. You want to sit in a classroom with a teacher who's guided previous toppers. You want the human accountability that comes from a mentor who knows your name.

This showed up clearly in willingness to pay. Students were happy to spend ₹500 to ₹1,000 per month on an AI tool — a useful supplement. But they were prepared to pay ₹20,000 to ₹1,00,000 to enroll in a human-driven coaching institute. The trust gap between AI and human guidance, measured in rupees, was 20-100x.

Where This Leaves Us

Nine months of building an AI tutor for UPSC gave us clarity on a few things:

A generic AI tutor has no moat. Students already have ChatGPT and Gemini. Another AI chatbot, no matter how well-tuned, isn't compelling enough to drive onboarding on its own.
Trust is the bottleneck, not technology. The AI needs a human anchor — a teacher, a brand, a face — to earn the right to be part of a student's preparation journey.
AI is a force multiplier, not a replacement. The winning model is AI + educators, where AI handles scale (practice, feedback, personalization) and humans handle trust (teaching, mentoring, accountability).
Motivation can't be engineered by product alone. External accountability structures — cohorts, mentors, deadlines — are necessary complements to even the best-designed learning tools.

These are lessons we've been carrying forward from our skilling experiments and now from test prep. The core insight remains: in education, the technology is the easy part. The hard part is understanding human behavior — what makes someone show up, stay engaged, and actually learn.

What we realized is that the teacher-student relationship isn't just a flow of information. It carries authority, trust, and the ability to push. When a human tutor you respect and trust tells you to push yourself — because they genuinely expect more from you — you do it. But when an AI says the same thing, you know these are empty words coming out of a chatbot. That intensity, that weight behind the words, is missing.

We're not giving up on AI in education. If anything, we're more convinced than ever that AI will transform how India learns. But the transformation won't look like an AI replacing a teacher. Teachers are not going anywhere — they need to evolve, but they remain irreplaceable for the parts of learning that demand trust, authority, and real human accountability. AI will play a critical role, but it'll look like every teacher having an AI that makes them 10x more effective — and every student having an AI that meets them exactly where they are.