HanziFlow

100% Local: No cloud APIs. All inference runs on-device via Ollama and MLX
Memory-Aware: Pipeline design for 16GB RAM constraint
Protocol-Based: Each tool follows a defined interface for composability

A comprehensive ecosystem of 6 interconnected tools designed for Mandarin Chinese study, all running locally on Apple Silicon.

Architecture

Writer — Character practice & dictionary. Built with Next.js, Hanzi Writer, CC-CEDICT.
Server — Backend API & data persistence. Built with Python, FastAPI, SQLite (WAL).
OCR — Screen text recognition. Built with PaddleOCR v5, Tkinter, Ollama.
Echo — Audio transcription pipeline. Built with mlx-whisper, yt-dlp, Ollama.
Talk — Real-time conversation. Built with mlx-whisper, Kokoro TTS, Sherpa-ONNX.
HSK Pipeline — Vocabulary ETL. Built with BeautifulSoup, SQLite. 11,545 entries processed.