anoman
Observability

Anomaly detection

Anoman watches your traffic for unusual behavior and raises anomalies — using a statistical rolling detector and a per-customer ML model — without ever blocking a request.

Never blocks

Detection is fire-and-forget

Anomaly detection runs alongside every request but never blocks it — scoring is fire-and-forget, so it adds no latency to the completion and can't fail a call. Anomalies are surfaced for you to review, acknowledge, or investigate after the fact.

Two engines

Statistical + ML, running together

Two detectors score every request in parallel and raise an anomaly when either flags:

Z-score / Welford (rolling window)

A rolling-window statistical detector that flags values far from your recent baseline, across four rules:

  • token_spike — a sudden jump in token consumption.
  • request_burst — an unusual burst of requests.
  • guardrail_storm — a spike in guardrail blocks.
  • cost_outlier — a cost far above your norm.

Isolation Forest (per-customer ML)

A per-customer machine-learning model trained nightly from your history, scoring each request against an 8-dimensional feature vector inline — catching multi-signal anomalies the single-rule detector can miss.

Both run per customer, so "normal" is learned from your own traffic, not a global average.

Anomalies API

List and acknowledge

Use GET /anoman/v1/anomalies to list anomalies with their severity and status, and POST /anoman/v1/anomalies/{id}/ack to acknowledge one. Anomalies can also fan out as email ops alerts.

# List detected anomalies (with severity + status)
curl "https://api.anoman.io/anoman/v1/anomalies?severity=high&status=open" \
  -H "Authorization: Bearer anm-sk-..."
Related: Anomalies also appear in real time on the Live Feed.

Catch the unusual

Open the Anomalies view to review, acknowledge, and investigate.