Deep dive
Read the full case study
The technical details, the tradeoffs, and the results — written down properly.
Read case study Back to projects
AI & Machine LearningEducationongoing
Kaiwa Coach
Local-first Japanese pronunciation coach with an on-device Rust speech engine: DTW scoring, pitch contours and cryptographically signed lesson packs.
2026 9 months 1 team members
179 passing
tests
107 signed packs
lessons
100% on-device
audio
988 KB packs
assets
Overview
Kaiwa Coach scores spoken Japanese entirely on-device. A Rust engine (called through Dart FFI) runs voice activity detection, YIN pitch tracking, mora segmentation and DTW alignment against native reference contours. Flutter handles the interface, encrypted local storage and sync, while Firebase acts as a control plane only — no audio ever leaves the phone.
The challenge
Pronunciation scoring usually means cloud speech recognition: costly per minute, privacy-sensitive, and useless offline. The app had to score mora timing, pitch and intelligibility locally, ship licensed content as verifiable packages, and be honest about what synthetic calibration can and cannot prove.
Our solution
Vendored Sudachi for offline kana analysis, aligned learner audio to 16 ms-hop reference contours with DTW, gated content mismatches, packed 107 lessons into signed versioned bundles (Ed25519 + SHA-256 with rollback), and shipped the engine behind Dart FFI after resolving iOS static-link symbol stripping.
Key features
DTW pronunciation scoring with per-mora timing, pitch and intelligibility
Pitch Studio F0 contour visualization
Signed, versioned lesson packs with rollback
Encrypted local database with Firestore sync
Spaced-repetition error profiles scoped per lesson
Content-mismatch gate that refuses to score the wrong exercise
Results & impact
On-device scoring — no audio leaves the device
107 signed lesson packs (17 MB) mined from Common Voice ja 26.0
74 engine + 43 Flutter + 62 pipeline tests green
App assets cut from 9.6 MB to 988 KB for contours-only packs
100% voicing agreement between content pipeline and engine pitch tracking
Documented calibration limits instead of claimed accuracy
Related Projects
Explore more projects similar to this one
AI & Machine Learning
2026
NihonGO! AI Japanese Tutor
NihonGO! pairs learners with 3D avatar characters for one-on-one Japanese conversation practice. A streaming tutor replies in pure Japanese while a separate analysis endpoint returns corrections, vocabulary and grammar notes. The product spans a Flutter app (mobile, web, desktop), a Hono/TypeScript backend on Supabase, and four LLM providers behind one interface — with an internal security audit completed before store submission.
FlutterDartRiverpod+9
View Project
AI & Machine Learning
2025
VietProStocks
VietProStocks combines Vietnamese market data, technical analysis and a retrieval-augmented LLM into a research assistant. The RAG stack — MiniLM embeddings, a FAISS index and a local llama.cpp model — is designed to run entirely on a Raspberry Pi 4 8GB, with an OpenAI-compatible backend as an escape hatch when more compute is available.
PythonFastAPISQLAlchemy+9
View Project
Mobile Development
2026
Mobile App Factory
Mobile App Factory is a Dart workspace where multiple products share every package without importing each other. Headless learning, mastery, curriculum and reward engines sit under a Flutter shell, with optional Flame games. A custom factory CLI validates products, content and architecture — 57 files checked against 9 rules — and a single verify script gates every release.
FlutterDartRiverpod+5
View Project
Start Your Project
Tell us about your project and we'll get back to you within 24 hours.
Reference Project:
Kaiwa Coach