AI Insights & Anomaly Detection

ProBeya's AI module provides statistical anomaly detection on KPI time series, trend forecasting, automated action suggestions, and AI-generated meeting summaries. These capabilities help operations teams catch deviations early, act on data-driven recommendations, and save time on administrative tasks.

Overview

Manufacturing and pharma operations generate vast amounts of KPI data daily. Without automated analysis, deviations go unnoticed until they become critical. ProBeya's AI features provide:

• Anomaly detection: Statistical analysis of KPI values to identify unusual deviations
• Trend forecasting: Time-series prediction showing where KPIs are heading
• Action suggestions: AI-generated recommendations based on anomaly patterns and historical resolution data
• Meeting summaries: Automatic generation of TIER meeting summaries with key decisions and action items
• Anomaly triage workflow: Full lifecycle from detection through acknowledgment to resolution or dismissal

Getting Started

Viewing Anomalies

1. Navigate to any KPI board with historical data.
2. Anomalies are automatically highlighted with markers on the chart.
3. Click an anomaly marker to see details: severity, expected range, actual value.
4. Acknowledge or investigate the anomaly from the detail panel.

Viewing Forecasts

1. Open a KPI chart with sufficient historical data (minimum 30 data points).
2. The forecast line extends beyond current data showing predicted values.
3. Confidence intervals are displayed as a shaded band.

Board-Level Insights Dashboard

1. Navigate to any board and open the AI Insights panel.
2. The dashboard shows a summary across all KPIs on the board:
   • Total anomalies detected (broken down by critical vs. warning)
   • Number of improving, declining, and stable KPIs
   • Per-KPI anomaly list and trend direction

How It Works

Anomaly Detection Algorithm

ProBeya uses pure statistical methods (z-score and moving average) rather than machine learning:

• Z-score detection: Computes a rolling mean and standard deviation over a configurable window (default: 7 periods). Flags values exceeding a configurable z-score threshold (default: 2.0 sigma).
• Moving average deviation: Calculates an exponentially weighted moving average and flags values deviating beyond a configurable threshold (default: 30%).
• Severity classification: Anomalies are classified based on the magnitude of deviation.

Configurable parameters per detection run:

Parameter	Default	Range	Description
windowSize	7	2-90	Rolling window for statistical calculations
stdDevMultiplier	2.0	0.5-5.0	Z-score threshold for anomaly flagging
deviationThreshold	0.3	0.05-1.0	Moving average deviation threshold

This approach was chosen over ML models because:

• It requires no training data or model maintenance
• It is fully deterministic and auditable (critical for GxP environments)
• It performs well on the typical KPI patterns in manufacturing (trend + seasonality + noise)
• It runs in real-time without GPU infrastructure

See ADR-023: AI Anomaly Detection Algorithm for the full decision record.

Anomaly Severity Levels

Severity	Criteria	Notification
Low	Minor deviation	No notification
Medium	Significant deviation (mapped from 0.3-0.5 deviation)	In-app notification to KPI owner
High	Large deviation (mapped from 0.5-1.0 deviation)	In-app notification to KPI owner
Critical	Extreme deviation (deviation >= 1.0)	In-app notification to KPI owner

Anomaly Triage Workflow

Persisted anomalies follow a lifecycle:

new --> acknowledged --> dismissed (false positive)
                    \-> resolved (addressed)

Status	Description
New	Just detected, awaiting operator review
Acknowledged	Operator has reviewed and is investigating
Dismissed	Operator determined it is a false positive or not actionable
Resolved	Root cause identified and corrective action taken

Trend Prediction

The predictKpiTrend and getForecast procedures use simple linear regression to predict future values:

• Fits a line through historical data points using least-squares regression
• Computes slope, intercept, and confidence (R-squared)
• Generates predicted values for a configurable number of future periods (default: 7, max: 90)
• Classifies the trend direction as improving, declining, or stable

For KPIs where direction is lower_is_better, the trend direction is automatically inverted: a positive slope (values increasing) is classified as "declining" because higher values are worse.

Notification Integration

When runAnomalyDetection is called, anomalies with severity of medium or higher trigger in-app notifications. If the KPI definition has an ownerId set, that user receives a notification with:

• The KPI name and anomaly severity
• The observed vs. expected value
• The percentage deviation
• The date of the anomalous reading

Tenant Isolation

All AI computations are scoped to the requesting organization:

• KPI data is filtered by organizationId before any analysis
• Anomaly records are stored with organizationId foreign key
• The verifyKpiOwnership helper ensures cross-tenant KPI access is impossible
• Meeting summaries are generated from tenant-scoped meeting data

API Reference

Procedure	Type	Description
aiInsights.detectKpiAnomalies	Query	Run real-time anomaly detection on a KPI's time series
aiInsights.predictKpiTrend	Query	Run linear regression trend prediction for a KPI
aiInsights.getInsights	Query	Combined anomalies + trends for all KPIs on a board
aiInsights.getAnomalies	Query	List persisted anomalies with optional status/severity filters
aiInsights.getForecast	Query	Get trend prediction with predicted future values
aiInsights.runAnomalyDetection	Mutation	Run detection, persist results, and trigger notifications
aiInsights.acknowledgeAnomaly	Mutation	Mark an anomaly as acknowledged
aiInsights.dismissAnomaly	Mutation	Mark an anomaly as dismissed (false positive)

Permissions

Action	Required Role
View anomalies and insights	Any board member
Run anomaly detection	Any board member
Acknowledge/dismiss anomalies	Any board member

Related Features

• KPI Boards — Define and track KPIs with thresholds and targets
• Benchmarking — Compare KPIs across sites
• TIER Meeting — Daily standup routines with AI summary generation