Tier 1 · Recall
Syntax & Primitives
Core formulas, protocol mechanics, mathematical constants, and fundamental definitions.
Tested: Scaled dot-product attention formula, MCP discovery transports, tokenization byte pairs.
Independent Research · Open Practice Hub
About CertScroll
CertScroll is a free, privacy-first study system for ten overlapping AI certifications. Clemens Wan built it at Wanessa Labs while preparing for the same certifications himself. It turns shared syllabus foundations into interactive concept graphs, diagnostic mock exams, hands-free audio drills, and a fast scroll feed.
Adaptive learning & auto-adjusting difficulty
CertScroll replaces passive flashcard flipping with an active, concept-anchored progression. Rather than asking random trivia, questions dynamically step up in cognitive difficulty as you master each concept.
Tier 2 · Discrimination
Architectural Trade-offs
Comparing competing approaches and choosing the optimal algorithm under real constraints.
Tested: LoRA vs QLoRA rank memory trade-offs, BM25 vs dense embeddings in hybrid search, DPO vs PPO.
Tier 3 · Application
Sizing & Workloads
Quantitative system calculations, hardware memory budgets, and production parameter tuning.
Tested: GPU VRAM allocation math (parameters × 2 bytes + KV cache), batch size step ratios, chunk overlap sizing.
Tier 4 · Diagnosis
Failure Analysis & Drift
Root-cause troubleshooting for production failures, model degradation, and adversarial exploits.
Tested: Vanishing gradient mitigation, Kolmogorov-Smirnov drift tests, indirect prompt injection defense.
The Auto-Adjusting Feedback Loop
How the scheduling engine personalizes your session in real time based on your responses:
✓ When answered correctly
The engine advances you up the cognitive ladder to higher difficulty application and diagnostic questions. Cards graduate from active review to long-term memory after consecutive mastery.
✕ When a question is missed
CertScroll pinpoints the misconception immediately with detailed distractor rationale and activates an inline "Reinforce concept →" drill with alternative questions on that topic until mastered.
Stratified Blueprint Mock Exams
Full exam simulations sample questions proportionally according to official syllabus domain weights stratified across all 4 cognitive tiers (~25% Recall, ~35% Discrimination, ~30% Application, ~10% Diagnosis). Every exam concludes with a concept mastery diagnostic breakdown and a 1-click "Start Adaptive Drill on Missed Concepts".
Continuous Shipping · Development Log
What's New in CertScroll
Chronological milestone log of every core subsystem, algorithm, and feature shipped across the project.
2Months active
16Logged milestones
2,050Questions in bank
10AI Cert Tracks
Shipping cadence by month
Counts reflect verified architectural milestones, algorithms, and subsystem releases.
September 20264 updates
Hands-Free Voice / Audio Study Mode & Review QueuefeaturePhase 5
Added Web Speech API TTS voice narration for question stems, multiple-choice options, and explanations with variable playback speeds (0.8x-1.5x) and hands-free auto-advance. Created human curation review queue tooling for certifying derived question cards.
clemenswan/certscrollService Worker Offline PWA & Calibrated Exam Readiness ScorefeaturePhase 4
Engineered offline-first PWA caching via Service Worker with automatic connectivity recovery. Built diagnostic exam readiness engine computing domain-weighted mastery, cognitive tier depth, and logistic pass probability gauge.
PR #50Interactive 3D Three.js Mind Map & Dedicated Mock Exams HubfeaturePhase 5
Launched WebGL Three.js interactive 3D constellation mind map on /bank/ with OrbitControls, fly-to camera animations, raycaster hover tooltips, and 3 layout modes. Built full /exams/ hub with domain-weighted blueprint exams.
PR #49Massive Bank Scaling to 2,050 Cards & 4-Tier Cognitive HierarchyfeaturePhase 4
Scaled question bank from 226 cards to 2,050 cards (2,005 questions + 45 concept cards) filling all 765 reachable matrix cells with 100% domain coverage across Recall, Discrimination, Application, and Diagnosis tiers.
PR #47
August 20268 updates
Mobile Slide Sizing Fix & /progress HUD Instrument RedesignimprovedPhase 3c
Resolved phone snap overflow on answered cards using :has(.card[data-answered]) slide growth and explicit Next Question control. Rebuilt /progress as a high-density instrument HUD with shaky-concept meters.
Shared Foundation Scope & Cross-Certification CompoundingfeaturePhase 3b
Introduced scope: "shared" on question cards with checkable rules. Expanded question bank coverage from 235 to 303 cells without authoring duplicate content, unlocking 75-165 shared foundation questions for unreadable guide tracks.
Citation Spine Gate & 205-Question Matrix FillfeaturePhase 3
Engineered guide spine leaf resolution gate checking sources by ordered subsequence containment. Populated 205 new derived questions spanning all 4 difficulty tiers.
Scenario Arcs, Auto-linking & Review Provenance TiersfeaturePhase 2
Added multi-step scenario arcs (e.g. loan-model), automatic concept term autolinking in stems, two-stage definition unlocking in ConceptSheet, and transparent source/review provenance on every card.
CSP Hydration Bug Resolved on Cloudflare PagesfixPhase 1b
Fixed Cloudflare Pages CSP header conflicts blocking Astro client hydration scripts. Switched to native Astro security CSP directives.
Core Scroll Feed Island & Dexie IndexedDB Local StoragelaunchPhase 1b
Built full-height snap scroll feed island with immediate single-tap grading, FSRS spaced repetition scheduling, ConceptSheet drawer, and local-first browser persistence.
Static Site Architecture & 10-Cert Syllabus CataloglaunchPhase 1a
Deployed Astro static site generation pipeline for 10 certification syllabi, concept detail pages, and static sitemaps at study.wanessalabs.com.
Phase 0 Standalone Scroll Snap SpikelaunchPhase 0
Constructed 167 KB zero-network single-file prototype verifying scroll-snap-type mandatory behavior, 100dvh mobile rendering, and touch response ergonomics.
Product Roadmap · Next Horizons
What's Ahead for CertScroll
Planned and shipped technical deliverables expanding AI certification readiness, offline resilience, and multimodal practice.
Hands-Free Voice & Audio Mode (Web Speech TTS)
Complete voice narration for questions, options, and explanations with speed controls (0.8x-1.5x), auto-advance, and keyboard shortcuts.
- Web Speech API integration with zero external API dependencies
- Auto-read on scroll / advance toggle for commuting & workout practice
- Instant audio playback for stem, options A-D, and rationale
Offline PWA & Calibrated Pass Probability Diagnostic
Offline Service Worker caching with shell precache and mathematical readiness engine modeling domain mastery against official 70% exam pass marks.
- Service Worker caching with background sync and offline status badge
- Radial SVG readiness gauge with 4 cognitive tier breakdowns
- Logistic calibrated pass probability model
3D Knowledge Galaxy & Stratified Blueprint Hub
WebGL Three.js 3D mind map linking 2,050 cards across 45 concepts, plus dedicated /exams/ blueprint simulator.
- Interactive 3D graph with 3 layout modes & smooth camera lerp fly-to
- Knowledge Map card accordions with live connected questions
- Configurable 15/25/50 question blueprint mock exams with timers
Human Curation & Content Review Queue
Interactive moderation and audit pipeline for graduating synthetic questions to human-verified status with leaf spine citation validation.
- Automated distractor balance, clarity, and leaf spine citation audit
- Interactive review and graduation tooling from derived to read tier
- Exportable curation verification log & audit reports
Agent-to-Agent Paid Question Bank (MCP & Bazantic API)
REST API and Model Context Protocol (MCP) server integration monetized via Bazantic.com, allowing autonomous AI agents to buy syllabus-cited questions, benchmark models, and run automated evaluation drills.
- High-performance REST / JSON API with OpenAPI 3.1 specification
- Model Context Protocol (MCP) server endpoints for question retrieval, grading & diagnostics
- Bazantic.com monetization integration for per-query agent micropayments
- Bounty revenue routing for human-verified (read tier) question creators
Zero-Knowledge Encrypted Multi-Device Sync
End-to-end encrypted backup and cross-device sync via WebRTC / QR transfer without requiring user accounts or storing plain text on servers.
- AES-GCM encrypted QR code import/export for instant phone-to-desktop transfer
- Optional WebRTC ephemeral peer-to-peer sync pairing
- Strict zero-account, zero-telemetry guarantee maintained
Multimodal Visual Architecture & Pipeline Puzzles
Interactive system architecture diagrams where learners diagnose bottlenecks, fix pipeline errors, and reorder transformer blocks.
- SVG pipeline canvas questions (e.g. RAG ingestion, LoRA injection points)
- Drag-and-drop architectural flow sequencing
- Touch-optimized visual diagnostic challenges
How the questions are sourced
Every question on this site is original, written from each provider's own published exam guide. None is copied, paraphrased, or reconstructed from a prep site, a braindump, or a real exam. Every card names its source, and a person reviews it before it ships.
What the links point to
Every resource on a concept or cert page links out to official training, an exam guide, or a public video. The site links to that material. It never mirrors it, and it never reproduces a provider's logo, badge, or trademark: certification names appear as plain text only.
Privacy & On-Device Guarantee
There is no account, and nothing to sign up for. Nothing about you leaves your device: your review history and progress are stored locally in IndexedDB, in your own browser, and this site has no way to see them. Cloudflare's cookieless analytics count page visits; there are no cookies, nothing that identifies you, and nothing that follows you to another site.
Because progress lives only on your device, it does not follow you to a different one. A phone and a laptop keep separate streaks. That is a deliberate tradeoff: the alternative is an account, and shipping something free that works today mattered more than syncing a streak across your devices.