Production anomaly analytics

Truweight production analytics

I built the anomaly and segmentation logic in production that separated genuine traffic risk from normal campaign noise, so the ops team could act on something real.

Employer / client: Truweight Wellness
Duration: Oct 2019 - Jan 2020
Project type: Analytics in production

Architecture

How the anomaly pipeline works

Anomaly detection pipeline: campaign and conversion signals pass through anomaly logic (threshold rules, model-based detection, precision-recall tuning), into a segmentation and routing gate, then into KPI reporting and support triage. Operator overrides loop back and retune the thresholds. — How campaign and conversion signals get split from real growth, routed to the right ops team, then reported, with overrides feeding back into the thresholds.
My anomaly-detection pipeline

Process flow

How I work the steps

01
before Signal capture
Campaign, chat, and conversion behaviour reviewed.
Marketing / Support
02
control Anomaly thresholding
Suspicious spikes split out from normal growth.
Data / Engineering
03
handoff Segmentation
Customer/support segments route likely issues earlier.
Data Analyst
04
after Operations response
Support triage and KPI review act on cleaner signals.
Support ops

How I worked the problem

Reframed bad traffic as a business decision: which source patterns, chat signals, and conversion drops should actually trigger action.
Worked with the engineers on the alert thresholds and segmentation logic so the model output drove campaign review instead of living as a clever notebook.
Tied the anomaly signals into support routing and KPI reporting so ops could react earlier with less manual triage.

Measured results

What I delivered

~15% reduction

Fraud-related campaign disruption reduction

After the anomaly tuning and validation discipline.

~10% sla

SLA adherence improvement

Support routing tightened up the response discipline.

Outcomes

Marketing pipelines got quieter. Spikes that used to ruin chat quality and lead scoring stopped reaching ops as a surprise.
Support triage routed itself. Agents got cleaner queues during peak load, and content saw which campaigns were actually working.