Anomaly Detection

Always-on watch over every object and every slice, before losses settle in.

Priori Sentry learns the normal shape of every metric in the context layer and alerts while a problem is still forming. This page explains how it watches, what it catches, and why the alerts are ones you can act on.

Priori Sentry

A weekly failure-rate chart spiking far above its normal shape — A real metric breaking shape: a weekly failure rate spiking to 6.8 percent against a baseline under half a percent, then settling back.

Why problems get found late

The monthly review is too late

Churn shows up in the review long after the users left. By the time the topline moves, the cohort or campaign that caused it is weeks old and the window to act is gone.

The average hides the problem

One sub-affiliate runs an 8% chargeback rate inside a portfolio averaging 0.6%. Dashboards show the top ten and the average; the 8% appears in neither.

Static thresholds cry wolf

A fixed alert fires every weekend because weekends are different, so someone mutes it. Muted alerts catch nothing.

What Sentry is

Sentry watches the metrics of the context layer: chargeback rates, conversion, message volume, refunds, per object and per slice. For each one it learns the normal shape, including seasonality, and scores what it sees against that shape.

It alerts on change, not on a number crossing a line. A weekend dip that happens every weekend is normal. The same dip on a Tuesday is not. And it stays quiet until the evidence is clear, because an alert feed the team trusts is the entire point.

How it watches

Coverage comes from the graph. Every object and every slice gets watched: each affiliate, each sub-ID, each MID, each campaign, thousands of series no dashboard would ever show. A new sub-ID is watched from the day it first appears, because the context layer knows it exists.

Alerts arrive where the team works, with the offending slice isolated and its history attached. When it matters, the on-call gets paged. Runner can pick the alert up from there and start the investigation.

What it catches

A card-testing run in its first minutes, showing as a burst of small authorizations on one MID. A chargeback cluster forming on a single sub-affiliate weeks before the processor letter. A campaign whose conversions stopped when a tracker broke. A drop in message volume from one geography after a deploy. The shape differs; the pattern is the same: small, specific, and invisible in the topline.

So what?

Hours instead of month-end. You cut the sub or fix the tracker while the damage is a day old, and the rest of the month's spend is saved.
An affiliate going quietly bad gets caught. Refund rates that creep for weeks never trip a static rule, and nobody rechecks a mid-list affiliate. Drift against its own baseline surfaces it while it is still cheap to cut.
Your chargeback ratio gets defended early. A cluster forming on one sub shows up weeks before the processor's letter, when dropping the sub still solves it.
Nobody has to watch the dashboards. A small team cannot staff a person to stare at charts. Sentry watches every affiliate, MID, and campaign at once, and stays quiet until the evidence is clear.

Book a demo All solutions