Case study
MeetingCopilot
Real-time bilingual meeting copilot for Windows
110 tests50 frontend + 60 Rust, all passing in CI
Live demo
What it is and what it solves
A Windows desktop app for Spanish–English meetings: live transcription, two-way text translation, private response suggestions and optional translated speech. The web demo shows the flow with fictitious data: meeting preparation, transcript, ES↔EN translation and suggested replies, with no AI and no recording.
Business version: MeetingCopilot →
Architecture
- Microphone → Rust core (Tauri)
- System audio → Rust core (Tauri)
- Rust core (Tauri) → Transcription
- Rust core (Tauri) → Translation
- Rust core (Tauri) → Copilot
- Rust core (Tauri) → React UI (IPC)
- Rust core (Tauri) → SQLite
Desktop app architecture. The web demo reproduces the interface with predefined data.
Three technical decisions
Separate microphone and system-audio channels
Who said what comes from the capture source rather than inferred speaker diarisation, so local input and remote playback never get mixed up.
Streaming with bounded queues
Audio is processed incrementally, while queue limits and stale-audio checks prevent an unlimited backlog: overload produces an explicit error instead of indefinitely delayed playback.
Rust owns credentials, networking and capture
React only calls specific Tauri commands; audio and the API key never pass through the frontend, and the production CSP restricts its connections.
Tests, with numbers
- Frontend: 50 passing tests across 17 files (UI, IPC contracts, transcription, translation, suggestions, history, voice and camera controls, static demo).
- Rust: 60 passing tests (audio processing, lifecycle and cancellation, simulated API exchanges, persistence, playback, caption-server security, OBS); 2 hardware tests run manually.
- ESLint, TypeScript, Vite build, rustfmt, Clippy and Windows installer packaging passing.
Verified
- Native desktop interface, React–Rust IPC and a generated Windows x64 installer (v0.1.0).
- Recorded microphone and system-audio capture checks.
- Voice processing, routing and OBS subtitles covered with synthetic audio and local mock servers.
Not claimed
- The web demo uses predefined data: no AI inference and no audio or video recording.
- The app keeps no audio recordings; saved transcripts live in local SQLite without app-level encryption.
Stack
Code and CI
Private repository · code walkthrough on request