landing page
selecting topic for duel
fetching questions in progress
chat funtionality for refining questions
duels history
overall leaderboar

Inspiration

The Moment No App Owned

You and your friend studied the same chapter. You both think you know it. There was no way to find out who was right — until exam results arrived too late to fix anything.

Every quiz platform — Kahoot, Quizizz, Quizlet — requires a host, a teacher, or a question bank you didn't write. None of them support the personal 1v1 moment:

"I know this better than you. Prove it."

SkillDuel was built to own that moment.

What it does

SkillDuel is an AI-powered 1v1 knowledge battle platform where anyone can instantly generate competitive quizzes from almost any source.

Paste lecture notes, a YouTube video, a PDF, a Wikipedia article, or even a Google Scholar paper. Gemini 2.5 Flash transforms the content into targeted multiple-choice questions within seconds.

Players send one challenge link to a friend, compete asynchronously, and receive a cinematic animated result reveal once both players finish.

After the duel, both players unlock a streaming AI coach that reviews every question in real time, explaining:

why an answer was correct,
where each player misunderstood concepts,
and what knowledge gaps they need to improve.

No teachers. No hosts. No pre-built question banks.

Just pure knowledge competition.

How we built it

The entire platform was built through multi-turn MeDo conversations across eight focused development sessions.

Session	Focus	What MeDo Generated
1 — Foundation	Auth, DB, landing page	React + Supabase auth, profile auto-creation, animated landing page
2 — Core Duel Flow	Duel creation, quiz, results	Duel builder, AI question generation, challenge links, timed quiz system, animated result card
3 — AI Reports & Stats	Performance reports, leaderboard	Knowledge gap analysis, AI performance summaries, Recharts dashboards
4 — Source Plugins	YouTube, PDF, Scholar, URL, Wikipedia	Transcript extraction, PDF parsing, article scraping, Scholar proxy integration
5 — Streaming AI Coach	SSE review system	Full SSE pipeline, Gemini streaming integration, eventsource-parser frontend architecture
6 — Public Duels	Discoverability & solo mode	Public duel feed, solo practice mode, score comparison system
7 — Game UI Overhaul	Game feel & immersion	Confetti victories, crown-drop animations, challenge overlays, TTS question narration
8 — Quiz UX Polish	Speed & responsiveness	Parallel audio prefetching, rapid transitions, dark-mode contrast optimization

The stack included:

React + TypeScript
Supabase
Supabase Edge Functions
Gemini 2.5 Flash
Server-Sent Events (SSE)
Recharts
Real-time subscriptions
Text-to-Speech APIs

Each session started with one high-level goal. MeDo generated the full-stack implementation — frontend components, backend functions, database migrations, and TypeScript types — through iterative multi-turn collaboration.

Challenges we ran into

Streaming state management

Rendering streaming AI responses word-by-word without causing excessive React re-renders was difficult. We solved this using refs combined with controlled state updates to keep the UI smooth during SSE token streams.

TTS latency

Question narration initially had noticeable loading delays. We optimized this by prefetching audio for the next question while the current one was still playing, reducing transition gaps to under 350ms.

Async multiplayer synchronization

Players finish duels at different times, so results had to remain synchronized and reveal instantly when the second player completed the match. Supabase Realtime subscriptions handled this elegantly.

Parsing multiple source formats

Each source type — YouTube transcripts, PDFs, websites, and academic papers — required different extraction and cleanup logic. Separate Edge Functions were built for every source pipeline to ensure stable AI question generation.

Accomplishments that we're proud of

Built a fully functional AI-powered multiplayer learning platform in only eight structured MeDo sessions
Created real-time streaming AI coaching with personalized explanations
Enabled quiz generation from nearly any knowledge source
Achieved fast, game-like UX with sub-350ms transitions
Designed a competitive asynchronous duel experience without requiring hosts or teachers
Delivered cinematic UI interactions that make learning feel competitive and social

One of our proudest achievements was the Streaming AI Review Coach.

After every duel, users can open "Review with AI" and receive a personalized coaching session streamed live word-by-word. The system analyzes exactly which concepts each player misunderstood and explains them conversationally instead of simply grading answers.

Building this normally requires:

SSE architecture,
frontend stream parsing,
incremental UI rendering,
AI streaming APIs,
and careful prompt engineering.

MeDo generated the entire pipeline in a single focused session.

What we learned

We learned that MeDo's multi-turn workflow becomes extremely powerful when each session is focused around a single clear objective.

Instead of over-specifying every detail upfront, the best workflow was:

Define one major feature goal
Let MeDo generate the full implementation
Refine specific edge cases iteratively through conversation

This approach dramatically accelerated development speed while still allowing deep customization.

We also learned how important UX responsiveness is for educational products. Small improvements like instant transitions, streaming feedback, and animated results significantly increased the feeling of competition and engagement.