Glyph AI is an AI-powered meeting assistant developed at Codeed Inc. that joins video calls, transcribes conversations in real-time, generates intelligent summaries, and creates structured action items reducing post-meeting documentation from 30 minutes to under 2 minutes. Built with Node.js for the real-time audio pipeline and Django for AI processing and data persistence.
Remote meetings generate enormous amounts of valuable information that is immediately lost buried in recordings nobody rewatches, notes only one person takes, and follow-ups that never get assigned. Glyph needed to be invisible during meetings and genuinely useful after them, without requiring any behavioral change from participants.
- Built a Node.js WebSocket server to handle real-time audio streaming from browser clients with sub-400ms latency.
- Integrated OpenAI Whisper for speech-to-text transcription with speaker diarization.
- Developed a Django backend for transcript processing, GPT-4 summary generation, and action item extraction.
- Designed a structured JSON output schema mapping cleanly to Notion, Slack, and Jira integrations.
- Built one-click export to Notion, Slack, and Jira so meeting insights live where teams already work.
The Meeting Problem
The average knowledge worker spends 31 hours per month in meetings. Of that time, an estimated 70% generates no documented output. Glyph was built around a single thesis: the value of a meeting is not the conversation it's the decisions made and the actions committed to. Everything else is noise that should be captured automatically.
Glyph integrates invisibly into existing video call workflows no new meeting tools required.
Meeting participant view and the post-meeting AI summary dashboard.
Real-Time Audio Pipeline
The Node.js WebSocket server receives audio chunks from the browser client at 100ms intervals, buffers them, and streams to OpenAI Whisper for transcription. Speaker diarization assigns each text segment to a meeting participant based on their audio track. The entire pipeline adds less than 400ms of perceptible latency transparent to participants.
GPT-4 Summary Architecture
Raw transcripts are chunked into 500-token segments with 100-token overlaps to preserve context across boundaries. A Django Celery task chain first extracts key decisions, then identifies action items with owners and deadlines, then generates a narrative summary. The structured output is validated against a Pydantic schema before storage, ensuring downstream integrations receive clean, consistent data.
NLP processing pipeline, structured action item extraction, and the integration connector system.
Integrations That Close the Loop
A meeting summary that lives in a separate tool is a summary nobody reads. Glyph's one-click integrations push action items directly into Jira as tickets, Notion as database entries, and Slack as threaded messages in the meeting's channel. The schema is designed so each integration receives exactly what it needs no reformatting, no copy-paste.
One-click Slack delivery and Notion export meeting insights delivered where your team already works.
Orderly The Second Project at Codeed
Alongside Glyph AI, I developed Orderly a food ordering application managing restaurant menus, real-time order queuing, and kitchen display integration. Both Glyph and Orderly share the same Django + Node.js architectural pattern, and working across both simultaneously gave me deep familiarity with building event-driven systems at production scale.
Glyph AI launched in closed beta with 50 teams at Codeed Inc.'s partner companies. Average meeting documentation time dropped from 30+ minutes to under 2 minutes. Teams that adopted in beta continued using it daily through the trial period.