Highlights

• 100% on-device AI — speech-to-text, a large language model, and neural text-to-speech all run locally; conversations, audio, and transcripts never leave the machine.
• Real-time voice pipeline — streaming microphone capture with RMS-energy / voice-activity endpointing and anti-hallucination transcript handling across MLX Whisper and live Apple Speech.
• Pluggable AI engine — one coaching core drives six interchangeable backends: Embedded MLX, llama.cpp (GGUF), LM Studio, Ollama, OpenAI, and Gemini.
• Low-latency neural TTS — Kokoro via MLX-Audio and native macOS voices, with sentence-pipelined synthesis and engine warmup so the coach starts speaking sooner.
• Full-stack desktop architecture — Electron UI plus a bundled, supervised Python/FastAPI backend with local SQLite persistence, packaged as a single installable app.
• Production security model — Ed25519-signed remote update/kill-switch, an anti-rollback monotonic clock, a tamper-evident local store, DNS-rebinding & loopback hardening, and token auth.
• Cross-platform release engineering — code-signed and notarized macOS builds via electron-builder with hardened Electron fuses, plus Windows/Linux build paths.
• Phone companion — a QR-paired LAN web app (PWA over TLS) that shares the same coach, backed by server-side cross-turn conversation memory.
• Privacy-first accounts & analytics — optional Supabase OAuth through a PKCE loopback handoff, an owner analytics console, and a signed offline licensing scaffold; telemetry stays content-free and opt-in.