AI Study Coach

Stop wasting time on generic practice tests. Generate targeted quizzes from your study materials, get intelligent feedback, and focus on actual weak spots—powered by Gemini AI on Cloud Run.

Comment 1

Inspiration

While preparing for competitive exams, I hit a wall: I knew my weak spots—fixed prepositions, specific historical events—but couldn't find targeted practice. Mock tests wasted hours on topics I'd already mastered.

Worse, when I finally found relevant questions, there was no feedback explaining why answers were wrong or how to improve.

I realized students everywhere face this inefficiency: generic practice materials that don't adapt to individual knowledge gaps. This led me to build an AI-powered study companion that generates personalized quizzes from any text and provides instant, contextual feedback—deployed as a scalable serverless application on Google Cloud Run.

What it does

AI Study Coach transforms any study material into personalized, adaptive quizzes powered by Google Gemini 2.5 Flash Thinking:

Smart Quiz Generation - Paste text from textbooks, notes, or articles → get targeted questions across 4 types:

  • Multiple choice
  • True/false
  • Fill-in-the-blank
  • Short answer

AI-Powered Evaluation - Goes beyond right/wrong grading with contextual understanding:

  • Identifies what was correct and what was missing
  • Provides detailed, encouraging feedback
  • Suggests specific improvements

Story Mode - Struggles with a concept? Get ELI5 explanations using analogies and simple language

Performance Analytics - After each quiz:

  • Identifies weak topics and strong areas
  • Detects misconceptions
  • Delivers personalized study recommendations

Serverless Architecture - Responsive web interface with frontend and backend scaling independently based on demand

How we built it

Architecture: Microservices on Google Cloud Run

Frontend Service

  • React single-page application built with Vite
  • Served via Nginx container
  • 512MB memory, 1 vCPU
  • Quiz state managed client-side using localStorage

Backend Service

  • Flask application with Gunicorn WSGI server
  • 2GB memory, 2 vCPUs (optimized for AI processing)
  • REST API with proper CORS configuration
  • Request-based autoscaling

Cost Optimization: Minimum instances set to zero for both services → zero idle costs while maintaining instant availability

AI Integration: Google Gemini 2.5 Flash Thinking

Designed 4 specialized prompts in AI Studio for different tasks:

  1. Quiz generation from text
  2. Subjective answer evaluation
  3. Story mode explanations
  4. Overall feedback generation

Prompt Engineering Best Practices:

  • JSON schema enforcement for structured output
  • Few-shot examples for consistent behavior
  • System instructions for quality standards
  • Saved in AI Studio with shareable links for transparency

Deployment Pipeline

  • Cloud Build automatically containerizes both services from Dockerfiles
  • Multi-stage builds for optimized frontend image size
  • Proper timeout configuration for AI processing
  • Environment variable management via Vite's import.meta.env

What we learned

Serverless architecture requires different thinking - Cloud Run's request-based autoscaling fundamentally changes design patterns compared to always-on servers. Splitting services by resource requirements (lightweight frontend vs. compute-heavy backend) optimizes both performance and cost.

Prompt engineering is a discipline - Production AI requires systematic iteration in AI Studio, not ad-hoc code tweaking. Structured prompts with JSON schemas, few-shot examples, and quality guidelines produce far more reliable outputs than natural language instructions alone.

Container runtime contracts matter - Cloud Run applications must:

  • Listen on the PORT environment variable
  • Respond to health checks within startup timeout
  • Handle graceful shutdown signals
  • Use production servers (gunicorn) not development servers

Microservices need careful integration - Communication across services requires attention to CORS, environment variables, and error propagation. Network failures must be handled gracefully with appropriate timeouts and user-facing messages.

What's next for AI Study Coach

Cross-device sync - Firestore integration for quiz history and progress tracking across devices using anonymous user IDs (no authentication required)

Analytics dashboard - Aggregate anonymized data to identify commonly missed topics and difficult question types, helping students discover areas others struggle with

Adaptive difficulty - Automatically adjust question complexity based on performance. Consistent success → harder questions; struggling → foundational review

Social features - Shareable quiz links, friend group leaderboards, and custom quiz challenges for increased engagement

Cloud Run Jobs - Batch processing for uploaded PDF textbooks and pre-generated quiz banks, demonstrating job execution model alongside current service architecture

Cost optimization - Intelligent caching for common question patterns, use lower-tier Gemini models for simple tasks while reserving Flash Thinking for complex evaluations

Share this project:

Updates