Loki

Your meetings stay private with Loki. Maximize group productivity while keeping sensitive information secure: capture, summarize, and remember your conversations without sending data to the cloud.

Comment

About Loki - Private AI Meeting Assistant

Inspiration

The idea for Loki was born out of a real privacy concern I had with existing meeting transcription tools. While solutions like Fireflies offer powerful features, they require sending sensitive meeting data to external servers. As someone who regularly participates in meetings containing proprietary information and confidential discussions, I found this trade-off unacceptable.

I was particularly concerned by recent trends where AI companies have been overstepping intellectual property boundaries to collect training data. The thought of my team's strategic discussions potentially becoming part of some future AI training dataset drove me to build a solution that keeps sensitive information where it belongs - on your device.

Data

How does it work?

Loki leverages a fully local AI pipeline that operates entirely on your device:

  1. Audio Capture: During a meeting, Loki captures audio from your system's microphone and/or speaker output.

  2. Real-time Transcription: The local Whisper model converts speech to text in real-time, displaying the transcription as the meeting progresses.

  3. Automated Processing: Once transcribed, the local Llama model analyzes the content to:

    • Generate concise meeting summaries
    • Extract key decisions and action items
    • Build a semantic index of the discussion

  4. Interactive Assistant: After the meeting, you can ask natural language questions about any part of the discussion. The local Llama model retrieves relevant information from the transcript and generates contextual answers.

  5. Secure Storage: All data remains encrypted on your device, with no information ever leaving your computer.

Why is Llama / On-Device AI a perfect fit for this project?

Using Llama and on-device AI is essential to Loki's core value proposition for several reasons:

  1. True Privacy: Financial institutions, healthcare providers, and legal firms handle extremely sensitive information that should never leave secure environments. On-device processing guarantees this privacy.

  2. Regulatory Compliance: Many industries have strict data handling requirements (GDPR, HIPAA, etc.). Local AI processing helps organizations comply with these regulations by keeping data within their control.

  3. IP Protection: Internal business discussions often contain trade secrets or proprietary information. On-device AI eliminates the risk of this data being used to train future AI models.

  4. Offline Capability: Loki works reliably without internet connectivity, making it suitable for secure facilities or locations with restricted network access.

  5. Cost Efficiency: For organizations with many meetings, the cost of API-based solutions can quickly become prohibitive. Local processing eliminates ongoing API fees.

What I learned

This project pushed me to deepen my understanding of several key areas:

  • Local AI deployment: Working with on-device AI models presents unique challenges compared to API-based solutions
  • Audio processing pipelines: Capturing, buffering, and processing real-time audio requires careful optimization
  • Electron architecture: Balancing performance across main and renderer processes while handling resource-intensive AI tasks
  • Privacy-first design: Building features that maintain utility without compromising data security

How I built it

Loki is built as an Electron desktop application using Vite for faster development and hot module replacement. The core architecture consists of:

  • Electron/Vite frontend: Creating a responsive UI that provides real-time feedback during meetings
  • Local Whisper model: For on-device speech-to-text transcription without sending audio to external services
  • Local Llama model: Running inference directly on the user's machine to generate summaries and answer questions about meeting content
  • SQLite database: Storing transcripts and derived data locally with encryption

The application captures audio directly from system sources, processes it through the local Whisper model, and stores the transcription. The Llama model then analyzes this text to generate summaries and create a knowledge base that can be queried after meetings.

Challenges I faced

Building Loki presented several significant technical hurdles:

  1. Audio transcription accuracy: Getting clean audio capture from multiple sources (microphone, system audio) and properly segmenting speech for Whisper proved challenging. I had to implement custom buffering and noise reduction to improve results.

  2. Llama.cpp integration: Having never worked with llama.cpp before, there was a steep learning curve to properly integrate and optimize the model for desktop use. Memory management was particularly tricky when dealing with longer meeting transcripts.

  3. Performance optimization: Running two AI models locally (Whisper and Llama) is computationally expensive. I had to carefully balance resource allocation and implement background processing to maintain a responsive UI.

  4. Cross-platform compatibility: Ensuring consistent audio capture and model performance across different operating systems required significant testing and platform-specific code paths.

Despite these challenges, I'm proud to have created a tool that provides powerful meeting intelligence while maintaining absolute privacy. Loki demonstrates that we don't need to compromise on data security to benefit from AI assistance.

Built With

Share this project:

Updates