I built a real-time collaborative CV annotation tool using CRDTs and WebRTC

Description and use

Sightline CRDT lets ML teams collaboratively audit and fix bounding boxes in real time without merge conflicts, using Yjs CRDTs over WebRTC, voice commands, and direct COCO/YOLOv8 export.

Problem

Machine learning and computer vision teams need to audit and fix training bounding boxes collaboratively without merge conflicts.

Outcome

The tool enables zero‑merge‑conflict collaborative annotation with voice control and instant export, improving audit efficiency.

Workflow

• Open the Sightline CRDT web app
• Select an image to annotate
• Invite teammates via the P2P link
• Use the UI or voice commands to adjust boxes
• Export annotations in COCO or YOLOv8 format

Evidence

Sightline CRDT, an open in-browser tool for machine learning and computer vision teams to collaboratively audit and fix training bounding boxes without merge conflicts.

Source

Powered by Yjs CRDTs over WebRTC so multiple auditors can inspect the same image simultaneously with zero merge conflicts.

Source

Voice Patching: Speak verbal commands like "Nudge box left" to correct coordinates and labels hands‑free.

Source

Full COCO and YOLOv8 export directly from browser memory.

Source