Offline-First AI Voice Assistant

An offline-first AI voice assistant PWA where all speech-to-text transcription runs locally in the browser via WebAssembly (Whisper.cpp compiled to WASM). Audio never leaves the device.

LLM inference happens through a secure Backend-for-Frontend (BFF) proxy that adds rate limiting and key management, keeping API keys server-side. Non-blocking inference runs in a dedicated Web Worker so the UI thread stays responsive.

Results:

• 80% reduction in UI freeze time vs baseline single-thread approach
• < 1.5s STT latency on modern hardware
• Zero audio data transmitted to any external server