We were
wrong
about the
problem.
EdStart founded April 2025. Three pivots later, the product that worked wasn't the one we set out to build. The audience we built along the way survived every iteration.
The original bet.
EdStart launched as a B2B AI-powered personalised learning platform for school students — adaptive tools sold into schools, calibrated to individual learning styles. First product: a math problem-solving chatbot that walked students through worked solutions step by step.
ChatGPT shipped the same "Study & Learn" mode three months after we did. Yes — we were ahead. It didn't matter.
Two problems killed the bet. Students wanted instant answers, not pedagogically clean
walk-throughs — and free answers with video explanations already existed on platforms
like DoubtNut for ~90% of textbook questions. Meanwhile schools couldn't absorb the
cost without pushing fees beyond ₹10,000/year, which parents
wouldn't pay for a tool they themselves never touched.
Four hypotheses.
Each phase is a stated hypothesis with a result tag. Pivots happened when the hypothesis didn't hold under contact with users, distribution, or unit economics.
- 01apr · may 2025B2B · math chatbot
"Step-by-step AI problem-solving improves outcomes and is valuable enough for institutions to pay."
Students wanted answers, not pedagogy. ~90% of textbook problems were already solved free on DoubtNut. Schools couldn't price it in without pushing fees past ₹10K/year.
FAILED - 02jun · jul 2025B2C · learning journeys
"Students will pay for personalised learning paths if guidance is clear and adaptive."
Pivoted to direct-to-student. Tested through LinkedIn (wrong audience), a 50-student WhatsApp community (mostly friends), and 1:1 tutoring. Distribution stayed weak; feedback stayed thin. Hypothesis weakened, not killed.
WEAKENED - 03aug · sep 2025B2C · JEE roadmap
"Structured roadmaps alone can drive paid conversion in competitive exam prep."
A 5-month JEE roadmap with 100 questions per chapter. A handful of students paid without trying the product — which said something uncomfortable about how edtech buying actually works. Killed the product anyway; it wasn't exciting to build.
UNSATISFYING - 04sep 2025 → nowB2C · ClutchPrep · focus app
"Solving focus and discipline is a more urgent pain than solving content."
The consistent pain in every student conversation was distraction from reels and short-form content — not absence of content. App-blocker MVP shipped; grew organically through the Instagram audience. Engagement validated; paid conversion stayed under 1%. Engineering ≠ monetisation in tertiary edtech.
READ_CLUTCHPREP_CASE_STUDY arrow_forwardPARTIAL ✓
Jugaad Jeetu.
Mid-June 2025 I launched an Instagram meme page targeted at Indian students. Not as marketing — as primary research. The audience grew. The product changed three times. The audience stayed.
students reachable in two taps.
Direct line into the audience EdStart was trying to serve. Comments, DMs, polls — every product hypothesis got tested with real students before a line of code shipped.
followers in the first 30 days.
impressions early on.
Most founders try to acquire users after launch. Acquiring the audience before is cheaper and lets every pivot stay close to the people you're building for.
The runtime.
One platform, three audiences. Next.js + Flask + MongoDB carried the same backbone across every pivot — only the prompts, payment flows, and product surfaces changed.
Pivot-proof stack.
Auth, sessions, payments, content storage stayed the same across all four pivots. Only the product surface changed.
Right tool, right job.
Claude for doubt resolution (conversational, accurate). OpenAI for bulk question generation (cheaper at scale).
IG → site → app.
Distribution is owned, not rented. Every product launch starts from a warm 20K-follower audience.
Three hard parts.
Adaptive calibration, role-scoped views, and the conversion funnel — the three places where the platform actually earns its keep.
Adaptive difficulty per kid.
Each student carries a rolling skill estimate per topic. After every attempt, we update
the estimate via a lightweight Elo-style rule and
sample the next question from a band that's ±0.4σ around the
estimate — hard enough to stretch, easy enough to land.
Streaks of correct answers push the band up. Wrong-streak after streak pulls it back. Per-topic skill, not per-grade — a kid can be at grade-9 trig and grade-7 fractions on the same dashboard.
Three views, one truth.
Same database, three audiences, three permission scopes. Students see their own attempt history. Parents see their child's progress and time-on-task — never the raw answers. Teachers see classroom rollups and gap analyses, but only for their batch.
Built as a row-level scoping layer in the API. Each
request carries a role + tenant, and every query gets rewritten with the appropriate
$match on the way in.
# request
GET /insights
authz: role=teacher · tenant=ace_school · uid=t_412
# rewritten
db.attempts.find({
tenant_id: "ace_school",
student_id: { $in: roster(t_412) },
submitted_at: { $gte: this_week() }
})
# student would get
db.attempts.find({
student_id: "s_998"
})
# parent would get
db.attempts.find({
student_id: "s_998",
$project: { answer_text: 0 } # strip raw answers
})Conversion is a funnel.
Trial → checkout → paid. Each step has a different reason it fails. Razorpay handles payment, but the leak before payment — abandoning the cart on the checkout screen — is bigger than the leak inside payment. Async email recovery fires 30 minutes after a cart is left, then 24 hours, then never.
Not aggressive. Two emails, then it's over. The cart that doesn't come back at hour 24 was never coming back at hour 72 either.
"Value creation around learning often matters less than how effectively a product is marketed."
Indian edtech is divided into three spending layers: schools / institutes (lakhs/year), after-school coaching (₹20–30K/year), and tools / books / online courses (₹5–20K/year). Engagement and conversion drop sharply as you go down. EdStart sat in the tertiary layer, which made distribution non-negotiable.
High-quality AI is expensive. To a student, AI isn't itself a differentiator — a well-designed question bank often beats it on price-to-value.
I didn't raise. I didn't compete head-on with large coaching brands on content. I didn't force AI into products where it wasn't earning its keep. That negative space ended up being the most useful set of decisions.
The journey fundamentally changed how I think about value, distribution, and monetisation in Indian edtech. ClutchPrep is what it taught us to build.
Read the full essay, or jump to what EdStart became.