Cogniloop

The AI tutor that refuses to give you the answer.

Live: https://cogniloop-vaibhav4046s-projects.vercel.app

Most AI study tools summarize content for you — and quietly destroy your understanding. Cogniloop does the opposite: it forces active recall through the Feynman technique, drills you with Socratic questions, evaluates your explanations, and adapts each round to your blind spots.

You paste a topic.
Cogniloop extracts the concepts.
Cogniloop asks. You explain.
Cogniloop grades. Cogniloop probes the gap.
Loop until mastered.

---

What it does (115+ shipped features)

Core loop

• Locked Socratic system prompt — refuses to give answers, forces explanation
• 5 question types: explain, apply, contrast, predict, trace
• Adaptive difficulty 1–5, scales with your performance
• 0–3 scoring rubric per round with verdict + gaps + strengths
• Concept state machine: weak → shaky → solid → mastered
• Live concept tracker sidebar with progress bars
• 8-round soft cap with smart "should-end" detection

3 study modes

• Chill — patient, hints generously, low starting difficulty
• Exam — strict grading, 90s timer per round, mid-range difficulty
• Expert — adversarial, first-principles, starts at difficulty 3

8 curriculum templates — pre-loaded concept maps, click-to-start

• JEE Main (Physics, Chemistry, Math — NCERT)
• NEET UG (Bio, Physics, Chemistry)
• GATE CSE (data structures, algorithms, OS, DBMS, computer networks, compiler theory, digital logic, engineering math — 8 standalone subjects)
• MCAT (Bio/Biochem, Chem/Phys, Psych/Soc, CARS — all 4 sections)
• AP Computer Science (CSA + CSP)
• ML Fundamentals (linear algebra, probability, optimization, neural nets, unsupervised learning, transformers, generative AI/LLMs, RL, MLOps)
• System Design (distributed systems, databases, APIs, scaling, microservices — interview-ready)
• Economics (micro, macro, behavioral, international, financial markets — AP, college, CFA Level 1)

Input / output

• Voice input via Web Speech API (Chrome/Edge)
• Read-aloud questions via Speech Synthesis
• LaTeX/KaTeX math rendering in questions, answers, evaluations
• Live math preview while typing
• In-session LaTeX math syntax panel — toggle with t for a quick reference card covering common symbols and expressions mid-answer; pairs with live preview for STEM-heavy sessions (JEE/NEET/GATE/MCAT)
• Markdown export of full session
• URL-encoded shareable session links — no backend, no leaks
• One-click copy button on the active session question card — paste directly into Notion, Anki, or any notes app; c keyboard shortcut
• One-click "Print / PDF" button in the coaching report — triggers the browser print dialog to save the session as a clean PDF (light theme, headers/sidebar hidden); p keyboard shortcut
• Per-round feedback card with strengths + gaps as bullet lists
• Collapsible round-review in coaching report — question, answer, score, and verdict per round
• Live word count while typing — color-coded amber / accent / green as explanation depth grows
• Question-type-aware answer placeholder — five coaching-voice prompts matched to explain / apply / contrast / predict / trace
• Question-type-aware "Stuck?" starters — one-click seed phrase opens the right frame for each question type (Feynman, concrete scenario, mechanism diff, cause-effect chain, step-by-step trace)
• Browser tab title shows the active study topic (e.g. "Backpropagation · Cogniloop")
• Live running average score in session header — e.g. "Round 3 / 8 · avg 2.5/3" — once the first round is graded
• Coaching report subtitle and markdown export show total words written across all rounds (e.g. "5 rounds · 12 min · avg 2.3/3 · 847 words written") — concrete measure of explanation depth

Persistence + accountability

• Daily streak counter (Duolingo-style)
• 90-day activity heatmap
• Lifetime stats: sessions, rounds, mastered concepts, avg score, study time
• Full session history with searchable list and Study again one-click
• Session search empty state offers "Study this topic now →" to launch a session directly on any unmatched search phrase
• localStorage-only — no accounts, no logs, no servers

LLM layer

• Multi-provider with automatic fallback
• Default: Pollinations.ai free public endpoint (no API key, no signup)
• Optional: Groq Llama 3.3 70B if GROQ_API_KEY env var is set
• Retry with exponential backoff
• Strict JSON mode + multi-strategy parser (raw, fenced, slice)
• Edge runtime on every route

UX details

• Sleek dark theme, Linear-inspired
• Geist font, KaTeX math styling
• Keyboard-first: ? shortcuts panel, g+g home, g+t g+h g+w g+s quick-nav, / to focus, Enter to start · ⌘/Ctrl+Enter to submit, 1/2/3 mode switch, e to end, m/l mic/read-aloud, c copy question, f focus answer, h Stuck? starter, t toggle math panel, n new session, d download report, p print PDF, s share link, j copy Feynman journal prompt, r re-drill weak concept / retry on error, v toggle round-by-round review in report, b browse curriculum templates
• Per-round difficulty dots, mode badges, question-type tags
• Scroll-progress reading bar on /templates, /history, /why — accent-colored 2px bar fills as you scroll
• Back-to-top button on long-scrolling pages — appears after 20% scroll depth, smooth-scrolls to top, instant jump when prefers-reduced-motion is set
• Sticky concept tracker on desktop, collapsible mobile
• Session resume on browser refresh
• Empty-state coaching card on the home page
• Score trend sparkline in coaching report — color-coded dots across all rounds
• Rotating coaching hints during start, grading, and report phases — 8 / 8 / 8 pool, brand voice throughout
• Previously-drilled dot on each curriculum topic button in /templates — accent dot when that topic exists in session history; aria-label appends "— previously drilled" for screen readers; zero network requests, invisible to first-time visitors
• Filter result in /templates shows N of 8 packs · M topics — see at a glance how much drill content the current query covers
• Mode-colored dot on session history cards in /history — green=Chill, amber=Exam, purple=Expert for instant visual scan without reading the label
• Actionable error tips in session error state — 429 rate-limit, 404 not-found, and timeout each show a specific fix (wait+retry, switch provider, use Groq) with a direct Settings link instead of the generic "retry" fallback

Pages

• / — landing with mode picker, examples, curriculum cards, feature grid
• /templates — full curriculum browser with topic-level filter
• /history — heatmap, streak, lifetime stats, session list
• /why — comparison table vs ChatGPT/Claude, 7-question FAQ, 6 use cases
• /study — the active session
• /settings — Groq API key, provider preference, voice language, TTS rate/pitch
• /shared/[token] — read-only view of any URL-encoded session

Why this beats wrapping ChatGPT yourself

	Cogniloop	ChatGPT / Claude
Default behavior	Asks you to explain	Answers immediately
Bypass-resistance	System prompt + UI lock	One re-prompt away
Concept progress map	Live, structured	None — buried in scrollback
Difficulty scaling	Explicit 1–5	Improvised tone
Coaching report	Markdown export, study plan, journal prompt	Loose prose if asked
Streak / accountability	Yes	No
Curriculum templates	8 packs, 100+ topics	Cold-start every chat
Privacy	Browser-only	Server-stored
Cost	$0 forever	$0 limited / $20+ pro

Stack

Layer	Choice
Framework	Next.js 16 App Router
UI	React 19 + Tailwind CSS v4
Language	TypeScript (strict)
Math	KaTeX
Runtime	Edge functions on every /api/* route
LLM	Pollinations.ai (default, no key) → Groq Llama 3.3 70B (fallback)
Persistence	localStorage (history, streak, in-flight session)
Hosting	Vercel

Architecture

app/
├── page.tsx                  → landing
├── study/page.tsx            → session container
├── templates/page.tsx        → curriculum browser
├── history/page.tsx          → streak, heatmap, sessions list
├── why/page.tsx              → comparison + FAQ + use cases
├── settings/page.tsx         → API key, provider, voice preferences
├── shared/[token]/page.tsx   → read-only session view
└── api/session/
    ├── start/route.ts        → extract concepts + first question
    ├── turn/route.ts         → evaluate + next question
    └── report/route.ts       → coaching report

components/
├── Landing.tsx               → home with input + mode picker + examples
├── Session.tsx               → state machine + question/answer/eval + report
├── TemplatesView.tsx         → curriculum browser with filter
├── HistoryView.tsx           → heatmap + sessions list
├── WhyView.tsx               → comparison + FAQ + use-cases
├── SharedView.tsx            → read-only shared session
├── SettingsView.tsx          → API key + voice settings UI
├── NavBar.tsx                → top nav with g+x shortcuts
├── ModePicker.tsx            → Chill/Exam/Expert
├── StatsPanel.tsx            → streak / sessions / rounds / avg score
├── VoiceInput.tsx            → Web Speech API mic
├── ReadAloud.tsx             → Speech Synthesis
├── ShortcutsModal.tsx        → ? panel
├── ShortcutsLoader.tsx       → lazy-loads ShortcutsModal out of the critical bundle
├── ScrollProgress.tsx        → 2 px accent reading-progress bar (/templates, /history, /why)
├── Toast.tsx                 → ephemeral error / success notifications
├── Math.tsx                  → KaTeX renderer with $...$ + $$...$$
└── Logo.tsx

lib/
├── types.ts                  → SessionState, Round, Concept, Evaluation
├── prompts.ts                → SYSTEM_CORE + EXTRACT/EVAL/REPORT + mode overlays
├── llm.ts                    → multi-provider chat() with retry + JSON parser
├── sessionLogic.ts           → runStart / runTurn / runReport (shared by edge + browser)
├── settings.ts               → UserSettings read/write + browserCallCtx() helper
├── modes.ts                  → Chill / Exam / Expert
├── curricula.ts              → 8 curriculum packs
├── hints.ts                  → loading / grading / ending hint copy
├── storage.ts                → history, streak, lifetime stats
├── share.ts                  → URL-safe base64 encode/decode
├── fetchRetry.ts             → fetch with exponential-backoff retry
├── rateLimit.ts              → edge-side rate-limit middleware
├── labels.ts                 → QTYPE_DESC / STRENGTH_DESC / QTYPE_PLACEHOLDER / STUCK_STARTERS / STRENGTH_COLOR
├── kbd.ts                    → isInTextField keyboard guard (shared across all pages)
├── useAutoExpand.ts          → auto-expanding textarea hook
└── useGlobalNav.ts           → g+x chord navigation shortcuts

Run locally

git clone https://github.com/vaibhav4046/cogniloop.git
cd cogniloop
npm install
npm run dev

Open http://localhost:3000 — no .env required.

Optional: set GROQ_API_KEY in .env.local for higher-quality / faster responses via Groq Llama 3.3 70B (free tier available at console.groq.com).

Deploy

vercel deploy --prod

Edge runtime is already configured per route.

Roadmap

• Spaced repetition queue across sessions (return weak concepts on a schedule)
• PDF / image OCR ingestion for textbook pages
• Classroom mode — teacher generates topic + dashboard, students drill
• Anki / Obsidian export
• Self-hostable LLM backend (Ollama)

Author

Built by Vaibhav Lalwani — github.com/vaibhav4046

License

MIT