ML-Powered Adaptive IQ Test

A developer built IQ Platform, a free adaptive cognitive assessment tool that calibrates question difficulty based on age, education, and occupation. The tool uses a multi-feature ML regression model for scoring and remembers previously seen questions across sessions, all without a backend API or database.

Tags: nextjs typescript machinelearning webdev Most online IQ tests are broken. They ask the same 20 questions to everyone — whether you're a 14-year-old high school student or a 35-year-old PhD. They don't adapt. They don't remember what you've already answered. And after two or three retakes, you've memorised the answers. The score stops meaning anything. I got frustrated enough to build my own. After a few months of work, I launched IQ Platform https://iq-platform-plum.vercel.app — a free, fully adaptive cognitive assessment tool that calibrates question difficulty to your age, education, and occupation, scores you using a multi-feature ML regression model, and remembers which questions you've already seen across sessions. Here's the complete technical breakdown of how I built it. No Auth. No database. No backend API. Everything meaningful happens on the client or at build time. Here's why standard online IQ tests fail: Standard test: User A: 14 years old, high school → gets "What is 30% of 200?" User B: 26 years old, PhD CS → gets "What is 30% of 200?" Both answer correctly. Both score the same. But the question told you nothing useful about User B. The solution is adaptive difficulty — selecting questions based on who is actually taking the test. This is how real psychometric assessments like the WAIS-IV work. They calibrate to the individual. The question bank covers six cognitive domains: | Domain | What it measures | |---|---| | Numerical | Arithmetic, algebra, ratios, percentages, logarithms | | Verbal | Vocabulary, analogies, anagrams, antonyms | | Pattern | Number series, letter sequences, figurate numbers | | Logical | Syllogisms, seating arrangements, propositional logic | | Memory | Sequence recall, list recognition, position memory | | Spatial | 3D geometry, rotations, Euler's formula, clock angles | Each question has a difficulty rating from 1 school level to 5 postgraduate/doctorate level , plus optional minAge , maxAge , and tags STEM, arts, language, general : export interface Question { id: string category: QuestionCategory difficulty: 1 | 2 | 3 | 4 | 5 minAge?: number maxAge?: number tags?: string // 'stem' | 'arts' | 'language' | 'general' text: string options: string correctIndex: number explanation: string timeLimit: number // seconds per question } A difficulty-5 Logical question looks like this: { id: 'l8', category: 'Logical', difficulty: 5, tags: 'stem' , text: 'Cards A, D, 4, 7. Which to flip to verify "vowel → even on other side"?', options: 'A and 4', 'A and 7', 'D and 4', 'A, D and 4' , correctIndex: 1, timeLimit: 40, explanation: 'Flip A check even and 7 check no vowel . Wason selection task.' } While a difficulty-1 version looks like: { id: 'l1', category: 'Logical', difficulty: 1, minAge: 10, maxAge: 16, text: 'All cats are animals. Whiskers is a cat. Therefore:', options: 'Whiskers is an animal', 'All animals are cats', ... , correctIndex: 0, timeLimit: 15, explanation: 'Simple syllogism: cat → animal' } When a user fills in their profile, the system computes a difficulty target based on three inputs: js export const EDU DIFFICULTY: Record<string, number = { school: 1.5, diploma: 2.2, undergraduate: 2.8, postgraduate: 3.5, doctorate: 4.2, } export function getDifficultyRange profile: UserProfile { let target = EDU DIFFICULTY profile.education ?? 2.5 // Age-based adjustment if profile.age <= 13 target = Math.min target, 1.5 else if profile.age <= 16 target = Math.min target, 2.2 else if profile.age <= 19 target = Math.min target, 3.0 else if profile.age = 30 && isStem profile.occupation { target = Math.max target, 3.2 } return { min: Math.max 1, Math.floor target - 1 , max: Math.min 5, Math.ceil target + 1.5 , target } } So a 22-year-old B.Tech CS student gets a difficulty window of 2–4 , while a 14-year-old high school student gets 1–2 . The InfoScreen even shows a live difficulty tier preview as the user fills out the form — "Advanced tier — challenging questions requiring deeper reasoning" — updating in real time. The isStem helper checks the occupation string for keywords: function isStem occ: string : boolean { return 'engineer', 'developer', 'programmer', 'data', 'scientist', 'computer', 'software', 'hardware', 'cs', 'iot', 'electronics', 'ai', 'ml', 'cyber', 'network' .some k = occ.toLowerCase .includes k } STEM occupations get Numerical, Pattern, and Logical questions boosted. Arts/language occupations get Verbal questions prioritised. This was the feature that prompted the whole project. After 2–3 retakes, users were memorising answers. The solution: track recently seen question IDs in localStorage and always serve unseen questions first. js const SEEN KEY = 'iq platform seen questions v2' const MAX SEEN = 80 // remember last 80 question IDs function getSeenIds : Set<string { try { const raw = localStorage.getItem SEEN KEY return raw ? new Set JSON.parse raw : new Set } catch { return new Set } } function markAsSeen ids: string : void { const existing = ...getSeenIds const updated = ...ids, ...existing .slice 0, MAX SEEN localStorage.setItem SEEN KEY, JSON.stringify updated } During selection, unseen questions are sorted to the front: js const unseen = pool .filter q = seenIds.has q.id .sort = Math.random - 0.5 const seen = pool .filter q = seenIds.has q.id .sort = Math.random - 0.5 // Always prefer unseen const prioritised = ...unseen, ...seen selected.push ...prioritised.slice 0, perCategory With 147 questions and 20 per test, a user can take 7+ tests before any question repeats — and when repeats do happen, they're shuffled into different positions and contexts. This is the most interesting part technically. Instead of just calculating a percentage, I built a 7-feature weighted linear regression model that maps raw performance to an IQ estimate. js const features = { accuracyScore, // proportion of correct answers speedScore, // based on average time per correct answer consistencyScore, // penalises random correct answers guessing difficultyWeighted, // harder correct answers worth more categoryBalance, // uniform performance spikey adaptiveScore, // penalises timed-out questions educationNorm, // calibrates for education baseline ageNorm, // peak window 16-35 = 1.0 } Speed scoring — not just "was it right" but "how quickly": js function calcSpeed answers: TestAnswer : number { const correct = answers.filter a = a.correct && a.timedOut const avg = correct.reduce s, a = s + a.timeSpent, 0 / correct.length if avg < 6 return 1.0 if avg < 10 return 0.9 if avg < 15 return 0.8 if avg < 20 return 0.7 if avg < 25 return 0.55 return 0.4 } Consistency scoring — detects guessing by measuring variance within each domain: js function calcConsistency answers: TestAnswer : number { const byCategory: Record<string, boolean = {} answers.forEach a = { if byCategory a.category byCategory a.category = byCategory a.category .push a.correct } let totalVariance = 0, categories = 0 for const cat in byCategory { const results = byCategory cat if results.length < 2 continue const mean = results.filter Boolean .length / results.length const variance = results.reduce s, r = s + Math.pow r ? 1 : 0 - mean, 2 , 0 / results.length totalVariance += variance categories++ } return clamp 1 - totalVariance / categories 1.5, 0.3, 1.0 } If you get 3 correct and 3 wrong alternating in the same category, that high variance lowers your consistency score — which reduces your final IQ estimate. js const WEIGHTS = { accuracy: 0.42, speed: 0.12, consistency: 0.10, difficulty: 0.22, balance: 0.08, adaptive: 0.06, } const rawLinear = WEIGHTS.accuracy accuracyScore + WEIGHTS.speed speedScore + WEIGHTS.consistency consistencyScore + WEIGHTS.difficulty difficultyWeighted + WEIGHTS.balance categoryBalance + WEIGHTS.adaptive adaptiveScore Raw score 0–1 maps to IQ 70–145 using a z-score approach with the normal distribution: js // Map raw score to z-score, scale by education and age norms const zScore = rawScore - 0.5 4.2 educationNorm ageNorm // IQ = mean + z SD mean=100, SD=15 const iqEstimate = Math.round clamp 100 + zScore 15, 70, 145 // 90% confidence interval const ciMargin = Math.round 8 + 1 - consistencyScore 6 const confidenceInterval = iqEstimate - ciMargin, iqEstimate + ciMargin A raw score of 0.5 maps to exactly IQ 100 average . A raw score of 1.0 maps to approximately IQ 135. Demographic norms adjust the z-score — a doctorate-level user educationNorm = 1.07 needs a higher raw performance to achieve the same IQ estimate as an undergraduate educationNorm = 1.0 , reflecting calibration to education-group baselines. The percentile is calculated using a proper normal distribution CDF approximated via the Abramowitz and Stegun error function: js function erf x: number : number { const sign = x = 0 ? 1 : -1; x = Math.abs x const a1=0.254829592, a2=-0.284496736, a3=1.421413741, a4=-1.453152027, a5=1.061405429, p=0.3275911 const t = 1/ 1+p x const y = 1- a5 t+a4 t +a3 t+a2 t+a1 t Math.exp -x x return sign y } const percentile = Math.round 0.5 1 + erf iqEstimate - 100 / 15 Math.sqrt 2 100 Session history is stored entirely in localStorage , keyed by a normalised username: export function normaliseUserKey name: string : string { return name.trim .toLowerCase .replace /\s+/g, ' ' } This means "Om", "om", and "OM" all map to the same history bucket — but "Om" and "Rahul" are completely separate. No login required. export function saveSession profile, result, totalQ, correct, avgTime, diffRange { const session: TestSession = { id: session ${Date.now } ${Math.random .toString 36 .slice 2, 7 } , timestamp: Date.now , userKey: normaliseUserKey profile.name , iqEstimate: result.iqEstimate, categoryScores: result.categoryScores, mlFeatures: result.mlFeatures, difficultyRange: diffRange, // ... } const existing = getAllSessions localStorage.setItem STORAGE KEY, JSON.stringify session, ...existing .slice 0, 100 } The History page fetches all sessions, groups them by userKey , and renders per-user filter tabs. The IQ trend chart Recharts LineChart shows your score trajectory across sessions. Domain averages accumulate lifetime category accuracy. The homepage hero is a fully custom animated neural network rendered on a <canvas element — no library, pure browser APIs. 38 nodes drift slowly across the canvas. Synaptic connections form between nodes within 160px of each other, with opacity proportional to distance. Signal particles spawn on random connections and travel along them with a glow effect. js function spawnSignal { const from = Math.floor Math.random nodes.length // Find nearest neighbour within 180px let best = -1, bestDist = Infinity nodes.forEach n, i = { if i === from return const d = Math.hypot n.x - nodes from .x, n.y - nodes from .y if d < 180 && d < bestDist { bestDist = d; best = i } } if best == -1 signals.push { fromIdx: from, toIdx: best, t: 0, ... } } // Signal particle rendering signals.forEach sig = { sig.t += sig.speed const from = nodes sig.fromIdx , to = nodes sig.toIdx const x = from.x + to.x - from.x sig.t const y = from.y + to.y - from.y sig.t // Radial gradient glow const grad = ctx.createRadialGradient x, y, 0, x, y, 8 grad.addColorStop 0, sig.color + 'dd' grad.addColorStop 1, sig.color + '00' ctx.fillStyle = grad ctx.arc x, y, 8, 0, Math.PI 2 ctx.fill } It runs at 60fps on modern hardware and falls back gracefully on mobile. 1. Defining "adaptive" is harder than it sounds. My first version just randomised questions. Then I added difficulty tiers. Then age ranges. Then occupation tags. Each layer felt necessary once I had the previous one. The question is always: what dimensions of the user actually matter for question selection? 2. localStorage is surprisingly capable for personal data. No backend, no auth, no GDPR nightmare. For a tool that's genuinely private by design, browser storage is the right call — not a compromise. 3. Weighted regression beats percentage scoring immediately. When I switched from correct/total 100 to the 7-feature model, the results suddenly felt more meaningful. A user who answers 12/20 quickly on hard questions scores very differently from one who answers 12/20 slowly on easy questions — as they should. 4. The canvas animation took longer than the ML engine. The neural network hero is purely cosmetic but it's what people notice first and share. Don't underestimate the value of a visually distinctive entry point. 5. People want to retake immediately. The most common feedback in the first week was "I want to try again." The anti-repeat system was added within 48 hours of launching. Build for the thing people actually do, not the thing you imagined they'd do. 🔗 Try it: https://iq-platform-plum.vercel.app https://iq-platform-plum.vercel.app Built during my final year B.Tech in Computer Science at COER University, Roorkee. The platform is free, has no login, and stores nothing on a server. Questions are selected fresh for each session based on your profile, and the ML engine shows you a full breakdown of every feature that contributed to your score. Would love your feedback — especially on the ML model weights and whether the adaptive selection feels right across different education levels. Drop a comment or connect on LinkedIn https://linkedin.com/in/omprakash-984540299 . If you found this useful, consider leaving a reaction — it helps other developers find it. 🦄