Curie: Realtime Voice Intelligence

Every day, knowledge workers lose hours to meeting overload. Over 70% of meetings are considered unproductive, costing enterprises billions annually. Meanwhile, over 42% of enterprise IT departments are actively banning traditional AI note-taker bots because external bots enter calls as awkward attendee participants, violate GDPR compliance, and store sensitive proprietary audio on third-party cloud servers. Curie solves this crisis by introducing a zero-bot, client-side ambient architecture. Instead of deploying an intrusive bot into your call, Curie captures digital tab audio from Google Meet, Zoom, or YouTube alongside your local microphone directly in the browser using the Web Audio API. Audio is downsampled in real time to 16kHz linear PCM and streamed to AssemblyAI's Realtime Speech-to-Text API (Universal-3.6 Pro) over an encrypted WebSocket with sub-300ms turn latency. During the call, an always-on-top Document Picture-in-Picture window displays autoscrolling subtitles and session controls above full-screen slides or IDEs without context switching. The moment the meeting ends, Curie dispatches punctuated transcripts to the AssemblyAI LLM Gateway (Qwen 3.5) and Google Gemini Flash to extract a structured executive summary, key topics, decisions, and interactive action items with one-click PDF export. All streaming uses single-use ephemeral tokens, and zero raw audio is ever stored on servers.