DEV Community: Ganga Ponnu

Building an AI Quiz Pipeline without using an AI API

Ganga Ponnu — Mon, 27 Jul 2026 06:57:39 +0000

Sometimes the most valuable software isn't another AI model—it's the workflow around it.

The internet is full of AI quiz generators. Upload some content - click a button - a quiz appears.

So why did I build something that doesn't generate quizzes directly?

The Problem

Gemini (and other LLMs) are already very good at reading textbook pages.

But asking an LLM:
"Generate a quiz from this chapter."
rarely gives you production-ready output.

Some common issues:

inconsistent formatting
incomplete coverage of the chapter
questions in random order
difficult to import into another application

For a teacher, this means spending more time cleaning the output than actually teaching.

My Approach

Instead of building another AI wrapper, I built a workflow.
Textbook Images
│
▼
Optimized Prompt
│
▼
Gemini
│
▼
Structured Quiz Data
│
▼
CSV File
│
▼
RandomTyms Quiz Player

The AI does what it's good at:

understanding content
extracting knowledge
generating questions

My application does what it is good at:

enforcing structure
creating a reusable quiz format
powering an interactive quiz interface

Why CSV?

This is probably the first question developers ask.

Why not just return JSON?
Why involve CSV at all?

Because my users aren't developers. They're teachers.

CSV has some surprisingly practical advantages.

Opens in Excel.
Can be edited without technical knowledge.
Easy to archive.
Easy to share.
Works offline.
Can be imported into the quiz player. A quiz becomes a file that belongs to the teacher—not another conversation trapped inside an AI chat.

Prompt Engineering Is Part of the Product

One thing I learned while building this is that prompt engineering isn't just an implementation detail.

It's product design.

Instead of expecting teachers to write prompts like:

Generate 40 multiple-choice questions covering every important concept...

they simply copy a pre-built prompt.

That prompt tells the AI to:

analyse every uploaded textbook page
cover all major concepts
include definitions and key facts
produce output in a predictable structure

The complexity stays hidden. The workflow stays simple.

Why I Didn't Use an AI API

At first glance, that seems like extra work. But there were reasons behind it.

1. Zero backend AI costs

No API billing.
No token management.
No rate limiting.

2. Users benefit from better models automatically

When Gemini improves, the workflow improves and I don't need to deploy a new backend.

3. Human review
Teachers naturally review the generated questions before importing them.

In education, that small review step is often a feature—not a bug.

Trade-offs

This approach isn't perfect.
There is a manual copy-paste step.
That's friction.

But it's also where teachers can fix mistakes, remove unsuitable questions, or add their own.

For educational content, I think that's a worthwhile trade-off.

Not every workflow needs to be fully automated.

How I Built an AI Quiz Generator, a PWA Hub, and an Animated Kids' Brand — All on One Phone

Ganga Ponnu — Thu, 23 Jul 2026 12:13:45 +0000

I don't have a studio. I don't have a team. What I have is a smartphone — and that single constraint ended up shaping every decision behind RandomTyms, an AI-animated Tamil-English educational brand built around two characters, Lokesh and Varsha, and now a small suite of free learning apps.

This post is about the "how" — not the brand story, but the actual technical build: a PWA hub, an AI-powered quiz generator, and the scrappy fixes that come with building mobile-only.

What I ended up shipping

The core of it is a Progressive Web App hosted on GitHub Pages: randomtyms.github.io. It bundles together:

AI Textbook Quiz Generator

— generates quizzes from textbook content using Google Gemini, with OMR-style answer bubbles and score-tiered feedback

Spinner Wheel

— a simple randomizer for classroom use
Multiplication Practice — quick-fire drills for kids

Symmetry Detective

— an interactive shape-recognition tool
All free. No sign-up, no ads, no tracking.

Building the PWA hub

The hub itself is a manifest + service worker setup, which sounds simple until you're debugging it entirely from a phone browser's dev tools.

A few specific things that took longer than expected:

Live content feeds: the Blog tab pulls from a live Blogger JSONP feed, and the Shorts tab runs through an RSS-to-JSON proxy — no dashboard, just feeds updating automatically.

YouTube thumbnails silently failing: this one took a while to track down. Shorts-style RSS links didn't match the video ID regex I was using, so thumbnails just wouldn't render for some videos. Fixed it by adding /shorts/ pattern matching plus a fallback to the guid field.

Building the quiz generator

This is the piece with the most moving parts. It's not just "call an API and show the output" — the widget needed:

Prompt engineering tuned so Gemini reliably hits a minimum question count (it would sometimes shortchange this)
CSV export handling — cleaning up BOM characters and CRLF line endings that were quietly corrupting downloads
OMR-style answer bubbles and score-tiered feedback, built as a reusable embeddable component (rt-quiz-embed.html) and a full standalone version (rt-quiz-full.html)

Basic security hardening on the widget itself

One unexpected debugging detour:

Google Safe Browsing flagged the blog at one point. After digging in, the likely trigger wasn't the quiz code — it was the instructional pattern

"copy this prompt → open Gemini → paste → send"

reading as suspicious to automated scanners. Worth knowing if you're building anything that walks users through copy-pasting into another AI tool.

Making the repo actually discoverable
Shipping the apps was only half the job — a live PWA with no polish around it just doesn't get found.

Recent cleanup included:
A`` proper README.md with screenshots of each app, not just a text description

MIT license added (code only — the characters and brand art stay separately owned)
GitHub topics (education, pwa, kids-learning, quiz-app) so it surfaces in topic search
A pinned profile repo and a real profile picture instead of the default identicon

Small things, but they add up to a repo that looks maintained rather than abandoned.

What's next

More apps are in active development on the hub. If you're curious about building AI-powered tools without a team or a dev environment beyond a phone, happy to answer questions in the comments — this whole thing has been one long lesson in working within constraints rather than around them.

🔗 Try the apps: randomtyms.github.io
📖 Follow the blog: lokeshvarsha.blogspot.com