Presentation AI

What if PowerPoint had a brain... and didn't need WiFi?


A teacher in rural Bihar types "Explain photosynthesis for Class 10" on her phone.

47 seconds later — without internet — she has a complete presentation with images, videos, quizzes, and notes. In Hindi.

That's not a demo. That's Tuesday.


The Problem Nobody's Solving

Here's a number that should make you angry:

Teachers spend 5+ hours every week making presentations. That's 260 hours a year — just dragging text boxes and searching for images.

Meanwhile:

  • 📉 Students forget 90% of slide content within 7 days
  • 🌐 600 million Indians have unreliable internet
  • 🗣️ 22 official languages — but most EdTech only speaks English
  • 🤖 AI tools hallucinate facts teachers can't afford to be wrong about

Existing solutions? Pretty slides with shallow content. Cloud-dependent. English-only. Slow.

We asked: What if we could give every teacher — even in a village with no WiFi — the power to create Harvard-quality interactive lessons in under a minute?


What Presentation AI Actually Does

You type a topic. You get a complete learning experience.

Input:  "Explain the water cycle for Class 6"

Output (in 47 seconds):
├──  10 beautiful slides (auto-layouted)
├──  Relevant images (AI-generated)
├──  Educational YouTube video (auto-found)
├──  Diagram of the water cycle (auto-created)
├──  Teacher notes ("Pause here, ask students...")
└──  5-question quiz (auto-generated)

All of this works OFFLINE. In 20+ languages. With zero hallucinations.

The "Wait, Really?" Features

What You'd Expect What We Built
Text slides Videos + Images + Charts + Quizzes in every deck
"Requires internet" Fully offline — 1.5GB model runs on any laptop
English only 20+ Indian languages — speak your slides into existence
Same content for everyone 16 teaching styles — school kids ≠ corporate training
2-3 minute wait First slide appears in 362ms
Images sometimes fail 7 backup sources — 99.9% success rate

The Insight That Changed Everything

Most AI presentation tools work like this:

Topic → GPT → Text → Slide

That's a fancy typewriter.

We built an AI teaching assistant that thinks like a teacher:

Topic 
  → Research the subject (web search or local textbooks)
  → Structure for learning (not just information dumping)
  → Generate visuals that aid memory
  → Find supporting videos
  → Create assessment questions
  → Write presenter notes
  → Deliver in student's language

Same input. Wildly different output.


How We Built It

The Speed Problem

Original: Generate slides one-by-one → 2+ minutes of waiting

Solution: Generate ALL slides simultaneously + stream tokens in real-time

Before: Slide 1... Slide 2... Slide 3... (waiting... waiting...)
After:  [All 10 slides generating in parallel, content streaming live]

Result: First content appears in 362 milliseconds. Users see progress, not spinners.


The Offline Problem

Challenge: A teacher in Jharkhand has 2G internet. Cloud AI is useless.

Solution: We fine-tuned Microsoft's Phi-2 model on NCERT textbooks, then compressed it:

Original: 5GB model (needs GPU)
    ↓ Fine-tuned on Indian curriculum
    ↓ Quantized to 4-bit precision
Final: 1.5GB model (runs on CPU!)

The magic: RAG architecture means the AI physically cannot hallucinate. It only knows what's in the verified textbooks.


The "Images Always Break" Problem

Reality: AI image APIs fail. A lot. Broken images ruin presentations.

Solution: 7-layer fallback system

Try Unsplash (free stock) 
  → fails? Try Pexels 
    → fails? Try Pollinations AI
      → fails? Try FLUX-schnell
        → fails? Try FLUX-pro
          → fails? Try FLUX-dev
            → fails? Generate entire slide as image

Result: 99.9% image success rate. We've never shipped a broken presentation.


The Language Problem

India has 22 official languages. Most EdTech ignores 21 of them.

Solution: Bhashini API integration

  • Speak your topic in Tamil → AI understands
  • Generate content in Tamil
  • Read aloud in Tamil for accessibility

20+ languages supported. Not translated — natively generated.


The Tech Under the Hood

Layer Tech Why
Frontend Next.js 14 + Plate.js Real-time streaming, rich editor
Cloud AI GPT-4o-mini Fast, cheap, streams well
Offline AI Fine-tuned Phi-2 + llama.cpp CPU inference, no internet
Knowledge FAISS + NCERT embeddings RAG = no hallucinations
Images 7-provider chain Reliability through redundancy
Charts Gemini 2.0 Flash Extracts data from text automatically
Languages Bhashini API Government-backed Indian language AI
Database PostgreSQL + Prisma Type-safe, battle-tested

Numbers We're Proud Of

Metric Value Why It Matters
First content 362ms Feels instant
Full presentation 47 seconds Coffee hasn't cooled
Offline model 1.5GB Fits on any laptop
Languages 20+ Real inclusion
Image success 99.9% Reliability matters
Teaching styles 16 Personalization works

What We Learned

  1. Perception > Reality — 362ms feels instant. 60 seconds feels forever. Streaming changed everything.

  2. Offline is a superpower — Not a limitation. For 3 billion people, it's the only option.

  3. Retrieval beats fine-tuning for facts — You can't fine-tune away hallucinations. RAG can.

  4. Teachers are underserved — The best teaching tools are expensive, English-only, and need fast internet. That excludes most of the world.


What's Next

Feature Status
Upload your own PDFs → AI turns them into presentations Building
Voice control → "Add a slide about climate change" Planned
Animation engine → Auto-animate based on content Planned
Export as video → Narrated presentations Planned
Real-time collaboration → Multiple teachers, one deck Planned

The Bottom Line


Built for teachers. Powered by AI. Works offline.

Because every educator deserves superpowers.


Built With

Share this project:

Updates