All case studies
AI · EdTech Shipped Aug 2025 · 6-week build

Madrah.com — an AI tutor that kids actually finish.

A personalised Islamic learning platform for families. We took it from PDF deck to launched product, in production with 8,400 active families across 14 countries.

The Challenge

Madrah had a clear vision — Islamic learning that kids love — but no shipped product, no AI infrastructure, and a launch date 8 weeks out. Two prior dev shops had quoted 18 months.

What we shipped

  • AI-personalised lesson engine
  • Adaptive quiz generation (Claude + RAG)
  • Voice narration in 5 languages
  • Parent dashboard + family progress tracking
  • Stripe billing + onboarding flow

Stack

  • Next.js 15 · TypeScript
  • Claude Sonnet 4 · OpenAI Whisper
  • Postgres · pgvector
  • Vercel · Cloudflare R2
  • Stripe · PostHog
4.1×Lesson completion rate
8.4kActive families (month 3)
42dFrom kickoff to launch
14Countries served

The brief: ship a learning product kids actually return to

Madrah's founders had spent six months interviewing parents. The signal was clear: kids weren't engaging with traditional Islamic learning apps — the content was static, the design felt institutional, and there was no real personalisation. Parents loved the idea but bounced after one session.

They came to us with a 90-page Notion doc, a Figma file, and a promise to investors that the MVP would be live by end of Q3. Two agencies had already passed on the timeline.

"We needed someone who could see the whole thing — design, AI, engineering, content — and just build it. Growvate was the first team that didn't try to talk us out of the deadline." — Aisha M., Founder, Madrah.com

Week 1: audit, prioritise, kill scope

We started with a 4-hour founder session and an internal audit of the brief. The Notion doc had 27 features. We cut it to 9 for v1, with a roadmap for the next 18 — measurable, sequenced, and tied to engagement metrics.

Three decisions saved the launch:

  1. AI was the product, not a feature. Every lesson, quiz and feedback loop runs through Claude with a tuned system prompt. Static content was banned from v1.
  2. Voice from day one. Kids learn faster when content is read to them. Whisper for input, ElevenLabs for output, both available in 5 languages.
  3. Parent dashboard as the retention engine. Parents who saw weekly progress reports stayed. Kids whose parents stayed, stayed.

Week 2–3: design system + RAG architecture

Two parallel sprints. The design team built a kid-friendly UI system (rounded shapes, generous spacing, motion-led feedback) while the AI team designed a RAG pipeline that could serve personalised content from a curated 12,000-document Islamic source library.

The hardest call was the model. Claude Sonnet 4 won — better at long-context reasoning over scriptural material than GPT-4o in our evals, and significantly cheaper per million tokens at the volume we projected.

Week 4–5: building the loop

This was the build week most agencies skip — the part where the AI feature isn't just working but actually good. We ran 400+ evaluation prompts through the lesson engine, tuned the personalisation logic, and built a feedback loop that gets stronger every time a kid completes a lesson.

By end of week 5, every part of the product was live in staging. Parent accounts, kid accounts, lessons, quizzes, progress tracking, billing, voice — all working end to end.

Week 6: launch, observe, tune

We soft-launched on a Tuesday to Madrah's existing waitlist of 1,200 families. By Friday, 642 had signed up. By the end of the month, 4,800. By month three, 8,400 active families across 14 countries.

The metric we obsessed over — lesson completion rate — landed at 4.1× the industry baseline for kids' learning apps. We didn't reach it on launch day. We tuned our way there over the first 60 days, with weekly prompt iteration and A/B tests on the parent dashboard.

What we'd do differently

Two things. First, we'd invest in evaluation infrastructure earlier — the first three weeks of post-launch tuning would have taken half the time with a proper eval suite from day one. Second, we'd have started multilingual on day one rather than week 4 — retrofitting i18n into a fast-moving product is the single most painful kind of refactor.

Both are now standard parts of our 30-day sprint.

"Six months of consulting turned into three weeks of shipping. Our retention numbers don't look like a v1. They look like a product we've been iterating on for two years."

— Aisha M. Founder, Madrah.com
Next case study →

Operator.io — an autonomous ops agent for SMBs

Read next case