Protecting Revenue Before It Is Lost
Sales Anomaly Detection
A Croatian frozen food distributor faced a structural revenue protection problem: customer churn was silent, slow, and invisible inside their ERP.
Rivermind built and deployed a real-time sales anomaly detection system directly within the client's VENIO ERP — 11 specialised detector modules, 18 distinct signal types, four years of invoice history.
"No cloud dependency. No data leaving the client environment."
Reactive, Manual Customer Monitoring
Sales representatives only became aware of an account's deterioration when the BI department surfaced a report — by which point the customer had typically already been lost.
The VENIO ERP contained every invoice and line item. But extracting insights required manual Excel exports nobody had bandwidth to run systematically.
What Changed When Rivermind Was Deployed
84% of problems predicted 2+ weeks before they happen
Every account monitored automatically, every day
94.7% of losses flagged in advance, revenue-weighted
CRITICAL / HIGH / MEDIUM / LOW tiers sorted by € at risk
16 product groups tracked per customer, category-level alerts
Five-layer seasonal awareness, HORECA winter closure handled
Multi-Detector Ensemble with Intelligent Scoring
The architectural decision: 11 specialised detector modules each capturing a different dimension of customer deterioration, combined through an intelligent scoring layer.
Detectors are deliberately simple — the scorer's job is to separate genuine multi-signal evidence from noise accumulation.
Validated Through Rigorous Backtesting
Click any metric to expand the detail. All metrics measured against the actionable customer set only.
The backtest validates €3.7M/year in at-risk revenue flagged with 94.7% coverage. With a 43-day median early warning window, the sales team gains enough lead time to intervene before relationships become unrecoverable.
What This Deployment Taught Us
What This Deployment Taught Us
Keep detectors simple and make the scorer smart — separation of concerns is the difference between a maintainable system and constant debugging.
Backtest-driven tuning beats domain intuition. Initial weights set by reasoning were significantly wrong.
Evaluate on the actionable set only — confirming churned customers inflates metrics with trivial true positives.
Seasonal markets need layered handling. One seasonal discount was not enough — five mechanisms were required.
Score distributions matter more than thresholds — understanding the bimodal pattern led directly to the most efficient fix.
