Laxcorp-Research/project-raven

What it solves

Project Raven is an open-source meeting copilot designed to provide real-time AI assistance and automated post-call documentation without relying on a hosted backend or cloud session sync. It solves the problem of capturing high-quality, separate audio streams for both the user and the other participants while maintaining privacy by keeping recordings and notes local.

How it works

Raven is an Electron desktop application that uses native helpers (Swift on macOS, Rust on Windows) to capture system audio and microphone input. It employs GStreamer for echo cancellation and a residual echo gate to ensure clean transcription. The app uses a "Bring Your Own Key" (BYOK) model, sending audio to Deepgram or AssemblyAI for transcription and prompts to Anthropic or OpenAI for AI assistance.

Key technical components include:

• Dual-stream capture: Separates "You" and "Them" into parallel streams.
• Local RAG: Uses Xenova/all-MiniLM-L6-v2 for on-device embedding and retrieval of uploaded documents.
• Stealth Overlay: Uses setContentProtection to remain invisible during screen shares on platforms like Zoom and Teams.
• Local Storage: All session history and transcripts are stored in a local SQLite database.

Who it’s for

It is for professionals and researchers who need an AI meeting assistant but require strict data privacy, local control over their meeting history, and the ability to use their own AI API providers.

Highlights

• Real-time AI Assist: An overlay that suggests what to say during calls, invisible to other participants.
• Privacy-First: No Raven accounts or hosted servers; all data stays in local SQLite.
• Automated Recaps: Generates structured action items, follow-up email drafts, and talk-ratio analysis.
• Cross-Meeting Search: An "Ask across all meetings" feature that uses local retrieval to answer questions based on entire meeting history.
• Meeting Auto-start: Detects active meeting windows (Zoom, Meet, Teams, Webex) to prompt for recording.