Built the entire Phase 1 pipeline from scratch: wake word detection (“Hey Jarvis” via openWakeWord), offline speech-to-text with Vosk Indian English model, Piper TTS for voice confirmation, and Notion sync. Hardest part was notion-client v3 removing databases.query() — rewrote the pull function. Also found a rogue C:.git tracking my entire hard drive. First live test worked end-to-end: wake word detected at score 0.69, thought transcribed, saved locally, pushed to Notion, confirmation spoken back. Then upgraded to session mode — one wake word keeps listening for multiple thoughts until silence or “that’s all”.
