🚀 From Paper to Pixels: How I Built an AI-Powered Exam Platform with Google Gemini

Naman Dhakad — Tue, 03 Mar 2026 05:53:05 +0000

Transforming PDF question papers into interactive CBT experiences - a journey through hackathons, AI, and building for millions of students

The Spark ✨

It was 2 AM in the hostel room. My friend was frustrated—again. He'd solved a 180-question JEE mock test on paper, only to realize he'd made a calculation error on question #45 that threw off his entire ranking. No instant results. No analytics. No way to know his weak areas.

"If only there was a way to just... upload the PDF and take it like the real exam."

That question changed everything.

🎯 What I Built: Rankify

Rankify is an AI-powered platform that transforms any PDF question paper into a fully interactive computer-based test (CBT). No more paper. No more manual checking. Just upload, practice, and analyze.

The Magic Flow ✨

graph LR
    A[📄 PDF Upload] --> B[🤖 AI Extraction]
    B --> C[👁️ Review & Edit]
    C --> D[⚙️ Configure Test]
    D --> E[🎮 Take the Test]
    E --> F[📊 Analyze Results]

Features That Actually Matter 🎮

Feature	What It Does	Why It Matters
AI Extraction	Gemini reads your PDF and extracts questions	No manual typing
Confidence Scores	Each question gets a 1-5 confidence rating	Know what to review
Diagram Detection	AI finds and maps diagrams to questions	Visual questions handled
Multi-PDF Batch	Process 10 PDFs at once	Save time
Offline Mode	Works without internet	Study anywhere

🧠 The Hackathon Journey

This wasn't a weekend project. It started at an MLH hackathon where I had 24 hours to prove the concept worked.

Day 1: The Chaos 🌪️

Me at 6 PM: "I'll use Gemini to extract questions from PDFs!"
Me at 8 PM: *Reading Gemini documentation*
Me at 10 PM: *Making API calls that fail*
Me at midnight: *First successful extraction! 🎉*
Me at 2 AM: *It works but the JSON is garbage*

Day 2: The Pivot 🔄

I realized the hard way: JSON output from LLMs is unpredictable. So I built:

Zero-tolerance prompts - Instructions so strict Gemini had no choice but to behave
Fallback chains - If one model fails, automatically try the next
Confidence scoring - A multi-factor algorithm to rate extraction quality

// The fallback chain that saved my hackathon
const modelFallback = [
  'gemini-2.5-flash',      // Best quality
  'gemini-2.5-flash-lite', // Quality + speed  
  'gemini-2.0-flash',      // Reliable backup
  'gemini-2.0-flash-lite'  // Last resort
]

The Result? 🏆

A working prototype that extracted 85 questions from a JEE sample paper in 3 minutes. Not perfect—but proof it was possible.

🔥 What I Learned

Technical Wins 💻

Gemini's Vision API is underutilized

Most people use it for images. But PDFs? That's where the magic happens for educational tech.
Prompt engineering > Fine-tuning

I spent zero dollars on training. Just better prompts.
Client-side AI is viable

Privacy-first architecture actually became a selling point.

The Hard Lessons 😵

Problem	Solution
Rate limits	Implemented exponential backoff + model fallback
PDF parsing issues	Added pdf.js preprocessing
JSON hallucinations	Built a validation layer with confidence scores
Cost tracking	Created real-time usage dashboard

Soft Skills That Stuck 🎯

MVP-first thinking: Ship fast, iterate faster
User empathy: Students don't care about AI—they care about results
Failure tolerance: 70% of my experiments failed. That's okay.

📝 My Google Gemini Feedback: The Good, The Bad, The Ugly

✅ What Worked Brilliantly

Multi-modal = Game changer 🎨
Processing text + images together meant diagrams weren't a separate problem.
Context caching 💾
Reduced costs by 60%. Essential for production apps.
Large context windows 📏
1M tokens = entire question papers in one go. Unreal.
JSON mode 📋
Structured output without post-processing. So good.

⚠️ Where Google Could Improve

PDF parsing is inconsistent
Rotated pages, watermarks, and scanned docs still cause headaches.
Rate limits are tight
For a free tier, it's restrictive. I had to build a queuing system.
Error messages are cryptic
"Internal error" tells me nothing. Help us help you!
Cost visibility
Real-time spend tracking would prevent surprises.

😬 The Ugly Truth

Sometimes Gemini hallucinates. It sees a question it can't parse and just... makes up options. That's why confidence scoring exists now.

🚀 Where I'm Going Next

The journey continues:

Phase 1: Cloud Run Deployment

gcloud run deploy rankify-ai \
  --source . \
  --region us-central1 \
  --no-allow-unauthenticated \
  --labels dev-tutorial=blog-devcommunity2026

Phase 2: Agent Development Kit (ADK)

Building specialized agents for different subjects:

Physics Agent 🧪
Chemistry Agent ⚗️
Math Agent 📐

Phase 3: A2A Protocol

Agent-to-agent communication for complex multi-step reasoning

💡 To Anyone Building with Gemini

Start with Vision - It's more mature than you think
Build for failure - Every API call can fail
Hide the AI - Users want results, not technology
Trust but verify - Confidence scores are your friend

🎤 Final Thoughts

Building Rankify taught me that AI isn't about replacing humans—it's about amplifying. A what we can do student in a small town with a cracked phone can now practice with the same quality as someone with expensive coaching.

That's what keeps me building.

Thanks for reading! If you want to try Rankify, check out the live demo. And if you're entering the Built with Google Gemini: Writing Challenge—good luck! 🎉

MLH #DEVChallenge #GoogleGemini #Vue #Nuxt #AI #WebDev

Forem: Naman Dhakad