How we run real-time AI deposition analysis with Deepgram + Claude

📰 Dev.to · John Mahoney

Learn how to run real-time AI deposition analysis with Deepgram and Claude, enabling fast and accurate transcript processing

advanced Published 21 Apr 2026

Action Steps

Set up a Deepgram Nova-3 account to handle live transcript generation
Create a Node.js WebSocket server to receive and process the transcripts
Integrate Claude Haiku-4-5 with a 12-key JSON schema to analyze the transcripts
Configure the system to handle WebSocket idle timeouts and audio pipelines on macOS
Test the system to ensure accurate and fast results, with a goal of ~4 seconds per segment

Who Needs to Know This

This solution benefits developers and attorneys working together on deposition analysis, as it streamlines the process and provides quick results

Key Insight

💡 The hardest parts of the project weren't the AI models, but rather handling WebSocket idle timeouts, audio pipelines on macOS, and prompt engineering