Skyool.ai: Adaptive Learning Engine

The problem

Students preparing for high-stakes entry exams mostly learn by rote: generic question banks, expensive academies and practice papers that never explain why an answer was wrong. Nobody tells a student exactly where their gap is, or what to do about it today. Skyool wanted a product that could, across six very different exams, on the mid-range phones its students actually use.

Our approach

We treated the learner model as the product, not the AI label. The core decision was to write the learning engine as plain, platform-agnostic TypeScript with no browser, server or UI dependencies, so the exact same code runs on the server, in the web portal and inside the mobile app, covered by automated tests. Every exam became data on top of that engine (a knowledge graph, an official blueprint, a trap catalogue and scoring rules), so adding a programme never means rewriting the product. We also designed the brand around the idea: the first "o" in the wordmark is the student, the second is the target, and the space between them is the gap Skyool closes.

How we delivered it

Brand and engine architecture

• Custom wordmark, symbol states (gap, closing, ready) and a shared design-token system
• Platform-agnostic TypeScript engine shared by server, web and app
• Knowledge graphs per programme: subject, topic, skill, concept

Learner model and planner

• Adaptive Baseline diagnostic, about six questions per subject
• Bayesian Knowledge Tracing with guess, slip and forgetting curve
• Daily Mission planner with spaced repetition and next best action

Mistake intelligence and AI tutor

• Ten mistake types detected from answer choice, timing and confidence
• Clue Bank, automatic recovery sessions and named exam traps
• Ustad, a Socratic tutor whose method lives in the engine, plus an IELTS writing and speaking studio

Platform, admin and Android

• Student portal and installable PWA with Normal, Mock and Beast training modes
• Admin control room: students, question bank workflow, exam intelligence, payments, report cards, activity log
• Containerised deployment, daily backups, 150+ automated tests and a hardened Android app

What we built

One engine runs a loop on every answer: Assess, Diagnose, Plan, Learn, Practise, Test, Analyse, Adapt. A short adaptive Baseline estimates ability per subject. A Bayesian Knowledge Tracing mastery model with difficulty-aware guess and slip rates and a forgetting curve moves each skill up a seven-step ladder, from exposure to exam ready. A planner builds a Daily Mission sized to the time the student has, weighing exam importance, mastery, recent mistakes, time since review and days to the exam, with SM-2 style spaced repetition. Every wrong option is written to reveal a specific misunderstanding, so the engine records why a student was wrong, not just that they were. Ustad, the AI tutor, follows a Socratic method that lives in the engine: it asks for the student’s reasoning, names the misconception, gives one hint at a time and never reveals the answer before two real attempts, and it keeps teaching even if the language model is unavailable.

Results

Skyool runs six exam programmes on one engine, with a student portal and installable PWA, a private admin control room that sees every state on the platform, report cards with QR verification, leaderboards and an Android app, all deployed from one container with daily backups. Every score and readiness figure is labelled as an estimate with a confidence value, because honest numbers are part of the product.

• 6 exam programmes, 328 skills and ~2,000 original questions running on one engine
• 227 catalogued exam traps and 10 mistake types the engine can tell apart
• 240+ universities mapped to their entry tests
• 150+ automated tests across engine and server, 70k+ lines of TypeScript
• Web platform, admin control room and Android app from one codebase and one API