MandarinFlow — Real-Time Adaptive Mandarin Memory System

Inspiration

Most Mandarin learning apps teach vocabulary as static lists optimized for exams like HSK, not for real-world comprehension. In practice, spoken Mandarin is dynamic, frequency-driven, and heavily context-dependent.

This project was inspired by a simple gap: being able to recognize beginner words, but still struggling to understand real conversations in videos, media, and everyday speech. The goal was to build a system that teaches Mandarin the way it is actually used — not the way it is tested.

What it does

This is a mobile-first Mandarin learning system that replaces traditional flashcards with an adaptive memory engine.

It continuously trains users on Chinese vocabulary across multiple recall directions:

Chinese → English English → Chinese Pinyin → Character Character → Pinyin

Instead of static review sessions, the system dynamically adjusts what to test next based on user performance.

Words that are difficult are resurfaced more frequently, while mastered words are spaced out over time to maximize long-term retention.

The system prioritizes high-frequency, real-world spoken Mandarin rather than HSK or textbook ordering.

How we built it

The app is built as a modular, mobile-first web application.

Core architecture includes:

A structured vocabulary database (Chinese, Pinyin, English, metadata) A unified dynamic review queue system for all learning activity A spaced repetition engine that updates scheduling based on correctness and recall performance A quiz system supporting multiple recall modes (translation, typing, recognition)

All vocabulary flows through a single adaptive learning loop to ensure every item is actively tested and re-tested based on memory strength.

Challenges we ran into

The most significant challenge was designing a correct spaced repetition system.

Early versions contained a critical flaw where only the first vocabulary item in a session was being repeatedly tested, while other items were only shown passively. This broke the learning loop and prevented true retention measurement.

Solving this required redesigning the system around a unified review queue where:

every vocabulary item is actively tested no item is skipped or passively displayed repetition is driven entirely by performance signals, not ordering

Another challenge was avoiding reliance on HSK-based vocabulary ordering, which does not reflect real-world spoken usage.

Accomplishments that we're proud of

We built a fully adaptive Mandarin learning engine that behaves like a memory system rather than a static flashcard app.

Key achievements:

A working spaced repetition engine with dynamic scheduling Multi-direction recall training (not just translation-based learning) A unified review queue that ensures full coverage of all vocabulary A system designed around real-world spoken Mandarin frequency instead of exam-based ordering

Most importantly, we created a learning loop that continuously adapts to user memory strength in real time.

What we learned

We learned that effective language learning systems are not primarily content problems — they are feedback loop design problems.

The structure of how knowledge is reinforced matters more than the vocabulary itself. A correctly designed adaptive system can outperform larger datasets or more polished UI.

We also learned that simplifying the architecture into a single unified learning loop is more powerful than splitting learning into multiple modes (study, quiz, review).

What's next for Chinese Simplified / Pinyin / English Language Mastery App

We plan to expand the system into a real-world Mandarin intelligence platform:

YouTube video vocabulary extraction for learning from real spoken content Song lyric learning mode with handling for poetic and non-standard grammar AI-generated contextual example sentences for deeper understanding Speech recognition for pronunciation training and feedback Cross-device sync and personalized vocabulary adaptation

The long-term goal is to evolve this into a system that learns Mandarin the same way humans naturally acquire language — through repeated exposure, correction, and adaptive reinforcement.

Built With

• adaptive-learning-engine
• apis
• cloud-services
• databases
• fastshot-ai
• frameworks
• github
• local-storage
• medo-ai-builder
• mobile-first
• platforms
• progressive-web-app-(pwa)
• react
• spaced-repetition-system-(srs)
• supabase
• tailwind-css
• typescript
• unified-review-queue
• vite
• youtube-transcript-api-(planned)