EducraAI

An adaptive AI tutor that uses Gemini 3's thinking models and voice capabilities to make learning stick, powered by cognitive science.

Comment

Inspiration

Traditional learning apps treat education like a transaction — you ask, AI answers, you forget. 90% of what you learn today is gone within a week (Ebbinghaus, 1885).

We wanted to build something that actually helps you retain knowledge using proven cognitive science:

  • Spaced repetition to combat forgetting
  • Active recall to strengthen memory
  • Feynman technique to reveal understanding gaps
  • Transparent reasoning to learn HOW to think

When Google announced Gemini 3 with its transparent thinking process and real-time voice capabilities, we saw the perfect opportunity to create a truly adaptive learning companion.

What it does

EducraAI is a full-stack cognitive learning platform that combines voice-first tutoring with research-backed memory techniques.

🎙️ Voice-First Tutoring

Have natural conversations with your AI tutor using Gemini Live API. Visual waveforms show when you're speaking and when AI is responding. The AI adapts its teaching style in real-time based on your confusion signals.

🧠 Transparent Reasoning

See Gemini 3's thinking process step-by-step. Unlike black-box AI, students understand how the AI reached its explanation — building metacognitive skills.

📝 Living Notes

AI automatically extracts concepts, insights, and key points as you learn. No more manual note-taking — focus on understanding while EducraAI captures everything.

🕸️ Knowledge Graph

Visualize your learning as an interactive graph. See which concepts you've mastered (green), which are in progress (yellow), and which need work (red). The graph updates in real-time as you learn.

📊 Episodic Memory

Every learning session is stored as an episode with linked concepts. The Mind Stream shows your complete learning journey — what you learned, when, and how well you retained it.

🧪 Adaptive Quizzing

AI generates personalized quizzes with semantic answer matching. No more losing points for phrasing — we understand meaning, not just keywords.

🔔 Proactive Interventions

Get intelligent spaced repetition reminders based on memory decay algorithms. EducraAI knows when you're about to forget something and prompts review at the optimal moment.

📈 Research-Backed Metrics

Track your progress with metrics grounded in cognitive science:

  • Active Recall Score — How well you retrieve information
  • Feynman Readiness — Can you explain it simply?
  • Review Streak — Consistency builds mastery

How we built it

Frontend (Next.js 15 + React 19)

  • TypeScript with strict mode
  • TailwindCSS + Framer Motion for polished animations
  • ReactFlow for knowledge graph visualization
  • Mermaid.js for AI-generated diagrams
  • Zustand for state management with localStorage persistence
  • Web Audio API for real-time voice waveforms

Backend (FastAPI + Python 3.11)

  • Modular agent architecture:
    • GeminiAgent — Thinking models with configurable levels
    • GeminiLiveAgent — Real-time bidirectional audio
    • EpisodicMemoryAgent — Learning episode storage
    • ProactiveEngine — Spaced repetition scheduling
    • ConsolidationAgent — Memory decay simulation
  • WebSocket support for Gemini Live streaming
  • Optional Neo4j Aura for knowledge graph persistence

AI (Gemini 3 Flash)

  • Thinking Models with configurable levels (MINIMAL → HIGH)
  • Gemini Live API for real-time voice tutoring
  • Google Search Grounding for factual accuracy
  • Structured Output for quiz/concept extraction

Infrastructure (Google Cloud)

  • Cloud Run for serverless deployment
  • Cloud Build for CI/CD
  • Artifact Registry for container images

Gemini 3 Features Used

Feature Implementation
Thinking Models Visible reasoning steps in chat, configurable thinking budget (MINIMAL/LOW/MEDIUM/HIGH)
Gemini Live API WebSocket-based real-time voice tutoring with bidirectional audio streaming
Google Search Grounding Factual answers with citation support for accuracy
Structured Output JSON extraction for quizzes, concepts, notes, and knowledge graph updates
System Instructions Custom tutor personality with emotional intelligence

Challenges we ran into

  1. Audio echo cancellation — Solved by pausing microphone recording while AI speaks and resuming with a buffer delay

  2. Live API configuration — Had to discover correct model names and single-modality constraint through experimentation

  3. Quiz answer matching — Built robust normalization (lowercase, punctuation removal, prefix stripping) with semantic matching

  4. Episodic memory alignment — Ensured user_id consistency across components so Mind Stream shows correct episodes

  5. Knowledge context injection — Pass student mastery data to both text and voice sessions for personalized teaching

Accomplishments we're proud of

✅ Research-backed learning — Every feature is grounded in cognitive science (Ebbinghaus, Roediger, Bjork, Feynman)

✅ Transparent AI — Users see Gemini's thinking process, demystifying how AI reasons

✅ Production-ready — Deployed on Google Cloud Run with proper infrastructure

✅ End-to-end integration — Voice, text, notes, quizzes, and memory all work together seamlessly

✅ Beautiful UX — Polished animations, responsive design, and intuitive interactions

What we learned

  • Gemini 3's thinking models are incredibly powerful for educational applications — students learn better when they see the reasoning process

  • Real-time voice requires careful audio state management (echo cancellation, recording pauses, queue-based streaming)

  • Cognitive science principles (spaced repetition, active recall) can be effectively translated into software features

  • The Gemini Live API's transcription capabilities enable powerful voice-to-text learning pipelines

What's next for EducraAI

📱 Mobile App — Push notifications for spaced repetition reminders at optimal review times

👥 Collaborative Learning — Study groups with shared knowledge graphs and peer teaching

🔗 Integrations — Export to Anki, sync with Google Docs, import from Notion/Obsidian

🌍 Multi-language — Support for non-English learners using Gemini's translation capabilities

🎯 Curriculum Mapping — Align knowledge graphs with academic standards (Common Core, AP, IB)

Built With

Share this project:

Updates