Project Snapshot
Instant quizzes, adaptive drills, and detailed explanations powered by Azure OpenAI to help learners improve faster.
GenAINext.jsAzureEducation
Skills Flexed
- Azure OpenAI prompt engineering with JSON schema responses
- Next.js 15 App Router UI + accessibility-first interactions
- Adaptive learning logic: diagnostics, drift control, and retry flows
- PromptOps harness for regression testing evaluation suites
Why I built it
People asked me for “one more mock test” way too often. They needed something faster than authoring a Google Form, smarter than rote flashcards, and private enough to run from a browser tab. Quizzaro became that sidekick: paste any topic, pick a difficulty, and the app spins up a fresh multiple-choice quiz with explanations, grading, and targeted follow-ups.
Experience walkthrough
- Hello, Targeted Quiz — learners describe the concept they want to drill (e.g., “Binomial theorem at JEE Advanced difficulty”). Quizzaro calls Azure OpenAI with a JSON-schema prompt, generates questions + distractors, and auto-grades with reasoning. Missed an item? Hit Practice similar to missed and it loops in variations focused on the weak spots.
- Help me prepare — coaches set a broader goal. The agent maps subtopics, runs a diagnostic quiz, clusters gaps, then suggests a drill plan with progress nudges.
- Zero friction — no signup, all state stays in the session, keyboard navigation everywhere, and graceful fallbacks if the LLM is momentarily rate-limited.

Under the hood
- Stack: Next.js App Router + TypeScript, client components for fast quiz flow, Tailwind-esque utility styles for responsiveness.
- LLM contract: Azure OpenAI
gpt-4o-miniwithresponse_format=json_schema, validated through Zod so malformed answers never reach the UI. - Adaptive engine: wrong answers feed into a micro state machine that chooses the next prompt template, keeps difficulty balanced, and avoids repeating identical questions.
- Safety rails: guardrail prompt blocks for toxicity, retry with lower temperature, and friendly error banners if Azure rate limits the request.
- Accessibility: semantic regions, ARIA-labelled radios, high-contrast themes, and support for Markdown + LaTeX so math-heavy questions render cleanly.

What it taught me
- Designing prompt ops workflows is as important as building UI. I created a CLI harness to regression-test new prompts before shipping them.
- Lean telemetry (time to generate, retry counts, accuracy improvements) helps spot when the model drifts or when topics need different scaffolding.
- Speed matters: caching Azure embeddings + tightening payloads brought quiz generation under three seconds for most topics.
Curious to try it? Launch the live demo here: archits-quiz-app.vercel.app.