AI Anomaly Detection: When Your Metrics Tell You Something's Wrong

Why Dashboards Are Not Enough

Dashboards show what happened. Anomaly detection tells you something unexpected is happening now.

Real cases: Stripe webhook broke (revenue -40% for 6 hours). Deploy doubled latency (conversion -25%). Campaign went viral (server costs +800%).

Detection Pipeline

Metrics --> TimescaleDB --> Detection Engine --> Alerts

Pick Critical Metrics

Focus on revenue-impacting KPIs: hourly revenue, conversion, error rate, DAU.

Detection Methods

Method	Best For	Complexity
Z-score	Stable metrics	Low
Moving average	Trending	Medium
Isolation Forest	Multi-dim	Medium
Prophet	Seasonal	High

Smart Alerts

Reduce alert fatigue with severity tiers, cooldowns, and AI root cause summaries.

Quick Start: Z-Score

import numpy as np

def detect_anomaly(values, new_value, threshold=3):
    mean = np.mean(values[-30:])
    std = np.std(values[-30:])
    z = (new_value - mean) / std if std > 0 else 0
    return abs(z) > threshold

Z-score catches 80% of anomalies with zero ML complexity.