Our platform processes over 10 million events per second. We need a system that can detect anomalies in real-time with minimal false positives.
Requirements:
- Sub-second detection latency
- Less than 1% false positive rate
- Ability to handle concept drift
- Scalable horizontally
- Support for multivariate time series
We welcome both statistical and ML-based approaches.
Discussion (2)
@alex_ml2 hours ago
I've been working on quantization-aware training that maintains 98% accuracy. Happy to share my approach.
@edge_dev5 hours ago
Have you considered structured pruning combined with knowledge distillation? We saw 3x speedup on ARM.