PortfolioCase study
TS-01Live pilot · Twilio sandbox
TeachSpark
Ready-to-use worksheets, on WhatsApp.
A WhatsApp bot that helps a time-poor Indian K–12 teacher use AI for real classroom work — a differentiated worksheet in about two minutes.
17
teachers joined the pilot on WhatsApp
Measuredas of 24 Aug 2026
first week, test handsets excluded
47%
activated (8 of 17)
Measuredas of 24 Aug 2026
12 of 17 onboarded first (71%)
37.5 min
median time saved
Self-reportedas of 24 Aug 2026
as reported by activated teachers
Evidence keyMeasuredSelf-reported
The teacher’s problem
Generic AI doesn’t transfer into a teacher’s real week.
Teachers want AI to give them time back, but what exists is generic, fragmented and disconnected from their subject, grade and board. It started with one real teacher: Tushar’s mother, who teaches Sanskrit.
When I’m overwhelmed by prep and grading, help me solve this week’s specific teaching task with AI, so I get real time back… without having to become a techie first.
The WhatsApp experience
One two-minute chat, from “where do I start?” to a worksheet.
No new app, no prompt-writing. The teacher answers a few questions in WhatsApp and gets a ready-to-use worksheet as a PDF, plus the reusable prompt behind it.
- Welcome + value
- Grade, subject, board
- 30-second micro-lesson
- Pick the topic
- Worksheet PDF + reusable prompt
- Minutes saved
- Share with a teacher
Product decisions
Capability, not dependency.
- Could have
- Do the task for her: a worksheet-vending service
- Chose
- Teach the reusable skill, then show the time saved
- Because
- Problem-led, applied, capability-building learning with impact feedback was the corner nobody occupied.
- Could have
- A standalone app with a new install and login
- Chose
- WhatsApp, where teachers already are
- Because
- Distribution, not a new destination — through a Twilio sandbox for the pilot cohort.
- Could have
- Reposition it as an AI-learning course
- Chose
- Lead with the worksheet; teach inside the product
- Because
- The mentor’s challenge: “A teacher buys a worksheet that is good enough to give to her students tomorrow.”
How it works
A pure state machine behind one chat.
- Teacher on WhatsApp
- Twilio webhook
- Express server
- Pure state machine(teacher, message, now) → next step
- ClaudeSonnet 5 default; Haiku 4.5 switchable
- PDF / DOCX render
- Back to WhatsApp
Activation funnel
A real first week: 17 teachers joined, 8 activated.
32 event types were instrumented, and Tushar’s own test handsets are excluded from every number. The pilot came in under the targets it set for itself.
- 72Landing views
- 17Signed up · 23.6%
- 17Joined on WhatsApp
- 12Onboarded · 71%
- 8Activated · 47%
- 5Question papers exported
3
referrals
Self-reportedas of 24 Aug 2026
1 of 4
re-engaged by a nudge
Measuredas of 24 Aug 2026
What isn’t measured
- Under target: joined 40–50 and activation ≥ 60% were the goals.
- D1 retention couldn’t be measured: the window was shorter than 24 hours.
- No teacher interviews were recorded and no LLM output-quality evals exist.
What I learned
Three lessons from the first week.
Sell the worksheet, not “AI”
Lead with the artifact a teacher can use tomorrow; deliver the learning inside it.
Green tests aren’t correctness
335 passing tests still shipped a city map that placed zero sign-ups: teachers typed “Bangalore” four ways.
Smaller, honest numbers
Excluding my own test handsets the day before submission made every number smaller — and true.
Sources
Everything here is backed by a real artifact.
- PRD
- Design
- Architecture
- Code
- Test run
- Feedback
- Post
- Deck
- Live data