Scenario · Health, wellness & education

"Duolingo for GCSEs": what it takes to build an exam prep app, and what to cut

The cost to build an app like Duolingo for exam prep: what a tuition centre needs (adaptive diagnostics, AI marking), the weekly plan and the cost table.

By Akash Bansal, founding engineerUpdated 4 November 20267 min read

Illustrative scenario. The founder in this story is a composite, not a real client. The process, costs and decisions are drawn from builds I have done; names and business details are invented to show how a build like this typically goes.

Who is the founder, and what did she want to build?

Amara is a composite, an invented founder drawn from the questions tuition-centre owners ask. She runs a GCSE and A-level tuition centre in south London with a few hundred students across maths, sciences and English. Every term she hears the same thing from parents: "Can't they just practise on an app, like Duolingo?" Her plan was an app her students use between sessions, which she could later sell to other centres.

The brief she arrived with was a Duolingo clone: streaks, hearts, leagues, a mascot, native iOS and Android apps, and "AI that knows what each student needs". An agency had quoted her a range that started in five figures and ended in six.

What does an exam prep app actually need from Duolingo?

It needs Duolingo's engine, not its game layer. The part worth copying is adaptivity: choosing the next exercise from what the system believes the learner knows. Duolingo describes its model, Birdbrain, as predicting how likely a learner is to get a given exercise right, then building each lesson from a large pool of candidate exercises. It ran A/B tests showing that learners using Birdbrain-tuned lessons learned more (Duolingo blog).

Exam prep differs from language learning in two ways that change the product:

• There is a fixed target. A GCSE has a published specification and a known paper structure. "What does this student know?" becomes "Which parts of the specification will cost this student marks?"
• Much of the score is written answers. Six-mark science explanations and English essays can't be marked as right or wrong. That marking is what eats teachers' evenings, and it is where AI saves the centre real time.

So the product became three things: a short adaptive diagnostic, a gap map against the specification, and fast marking of written answers with a teacher in control. Streaks stayed. Leagues, hearts and the mascot didn't.

Why a founding engineer, and what's the proof?

She chose one senior engineer over the agency because the hard part was the logic (diagnostics and marking), not screen count. She also wanted to own the code before selling it to other centres. I can show the relevant work, and say exactly what state it is in:

• Brahmo (shipped) grades long-form exam answers. It is an event-driven, multi-LLM pipeline: Gemini OCR reads the answer, semantic matching lines it up against the expected points, and Claude scores it on a weighted rubric. A human expert can override any score. This is the marking engine Amara needed.
• Testing Engine (in development, not launched) is my AI precision-testing engine for competitive-exam prep: adaptive diagnostics, a gap map against the exam blueprint, and a model of the probability of passing, for both B2C and B2B2C. It is still being built, so I have no results to quote. The design thinking applies directly to Amara's diagnostic.

If all you need is a question bank with quizzes, you don't need a custom build. Off-the-shelf quiz and LMS tools do that well. Custom pays off when the diagnostic and the marking are what you sell.

What got cut from v1?

Week by week: how the build went

1. Week 1: scope and blueprint. A design doc, the fixed quote, and the exam specification as a structured map of topics, sub-topics and the paper each one sits on. Everything else hangs off this map.
2. Week 2: question bank import. Amara's own questions moved from Word documents into a tagged bank, with each question tied to specification points and a difficulty level.
3. Week 3: diagnostic v1. A short adaptive test. Each answer updates an estimate of mastery per topic, and the next question targets the topic the system is least sure about. A simple rating model came first, with a calibrated model (the item response theory approach behind adaptive tests like the GMAT) deferred until there was enough answer data.
4. Week 4: the gap map. A student view and a parent view showing which specification areas are strong, weak or untested, plus the next three things to practise.
5. Weeks 5–6: written-answer marking. The Brahmo pattern: a photo or typed answer goes through OCR, gets matched to mark-scheme points, and is scored on a weighted rubric with feedback a student can act on. Any mark under a confidence threshold goes to a teacher.
6. Week 7: teacher console and safeguarding. Teachers review flagged marks and override them. Overrides are logged and feed back into rubric fixes. High-privacy defaults, minimal data collection, and no public profiles.
7. Week 8: pilot. One year group, one subject, for half a term.

How much does it cost to build an exam prep app like Duolingo?

Agency guides put a full Duolingo-style app at roughly $50,000–$150,000, with a lean MVP at $25,000–$55,000. Appinventiv quotes $25,000–$100,000 for one platform (both are vendor sources). Those figures price the whole game layer. Amara's version is narrower. To see what your own feature list adds up to, the app cost calculator gives a rough range in dollars and weeks.

At around a penny an answer, a student submitting five written answers a week costs a few dollars a year to mark. That is what makes "AI marking included" a price she can offer other centres.

What went wrong, and what surprised her?

• Content was the bottleneck, not code. Tagging hundreds of questions against the specification took her teachers longer than the engineering took. Build the import and tagging tools early.
• Past-paper licensing. She had planned to load exam boards' past papers. Their use is governed by the boards' copyright terms, so the safe v1 used her centre's own questions, with past papers only where the terms allowed.
• Teachers wanted to see why. A mark of 4 out of 6 wasn't enough. "You explained osmosis but didn't mention the partially permeable membrane" was. Point-by-point feedback against the mark scheme earned the teachers' trust.
• Handwriting varies a lot. Poor photos produced poor OCR. A "retake photo" prompt and typed answers as a fallback fixed most of it.

What does the law ask of an edtech app for under-18s?

In the UK, the ICO's Age Appropriate Design Code covers online services likely to be accessed by children, not only those aimed at them. Its 15 standards include high-privacy defaults, data minimisation, limits on profiling, and geolocation switched off by default. An adaptive engine is a form of profiling, so the design doc recorded what is profiled, why, and who sees it. Treat this as a design input and get specific advice. I'm an engineer, not a lawyer.

What did launch look like?

A half-term pilot with one year group in one subject. Students took the diagnostic in a lesson, practised at home, and submitted written answers. Teachers marked only the flagged answers. The useful early signal wasn't a grade improvement (half a term is too short) but whether students came back between sessions, and how many AI marks teachers overrode. Both were tracked from day one.

Checklist: building an exam prep app like Duolingo

1. Copy the engine (adaptivity), not the game layer.
2. Map the exam specification first. It is your data model.
3. Start with a simple mastery model. Calibrate once you have answer data.
4. Use AI marking against a mark scheme, with a confidence threshold and teacher override.
5. Build a PWA first. Go native after daily use is proven.
6. Budget teacher time for writing and tagging questions. It is the real cost.
7. Design for the Children's Code from day one if under-18s will use it.
8. Build tenants into the data model if you plan to sell to other centres.

If you run a tuition centre or an exam-prep business and the marking or the diagnostic is the product, this is the kind of build I do, from a fixed quote with weekly demos, and you own the code. Here's how I work. For the line-by-line cost logic behind these numbers, see what an MVP costs in 2026. For how the marking stays honest, read multi-LLM scoring.