Real-Time AI Data Platform

The operations team was monitoring critical business metrics through a dashboard that polled a REST API every 30 seconds. When something broke, they learned about it minutes later, sometimes through customer complaints. They needed real-time streams, anomaly detection that could tell a genuine outage from a normal traffic spike, and connections that stayed alive for thousands of concurrent operators across three time zones. Two previous vendors had quoted 'six months and a dedicated ops team' for this work.

We built a real-time pipeline using Django Channels for WebSocket connections. The authentication, permissions, and domain logic already lived in Django, so putting the real-time layer inside the same framework cut three weeks of glue code. The React frontend uses Socket.io for persistent bi-directional communication. A custom anomaly-detection model runs as a separate Celery worker, pulling from the same event stream without blocking message delivery. When the model flags something, alerts route to email and Slack within seconds, and include the last 60 seconds of context so the on-call engineer doesn't have to go hunting.

• Sustains 5,000+ concurrent WebSocket connections with sub-100ms message delivery
• AI anomaly detection catches 92% of data irregularities before a human would review them
• Incident response time fell from 15 minutes to under 30 seconds
• Customer-reported outages dropped 78% in the first quarter post-launch

Django Channels over a standalone WebSocket server meant we kept one auth system, one permission model, one deployment. Redis is doing two jobs: the channel layer for message routing and the cache for frequently accessed data. The anomaly model runs in isolated Celery workers so a model hiccup can't back up the real-time queue. Horizontal scaling is per-service. We add Channels pods when connection count grows and Celery pods when event volume grows, and the two never compete for the same resources.