Everything you need for
anomaly detection

A modern, self-supervised engine that drops in where Azure left off.

100% Compatible

Azure API Compatibility

Canary Edge accepts the same JSON format and exposes the same endpoints as Azure Anomaly Detector. Your existing client code works without modification -- change the base URL and you are done.

  • /detect/entire -- batch detection over the full series
  • /detect/last -- streaming detection on the latest point
  • Identical request and response JSON schema -- zero code changes

Field Mapping

Azure Field Canary Field
seriesseries
series[].timestampseries[].timestamp
series[].valueseries[].value
granularitygranularity
maxAnomalyRatiomaxAnomalyRatio
sensitivitysensitivity
customIntervalcustomInterval
periodperiod
isAnomalyisAnomaly
isPositiveAnomalyisPositiveAnomaly
isNegativeAnomalyisNegativeAnomaly
expectedValuesexpectedValues
upperMarginsupperMargins
lowerMarginslowerMargins

Encoder

Compress raw time-series into latent representations

Predictor

Forecast expected latent state from context window

Energy Scorer

Compute reconstruction energy as anomaly signal

Regime Classifier

Map energy to operational regime via z-score thresholds

LeWM World Model

Canary Edge is powered by LeWM -- a Learned World Model inspired by Joint-Embedding Predictive Architecture (JEPA). It learns the normal dynamics of your system without any labeled anomaly data.

Self-supervised learning

No labels required. The model learns normal behavior by predicting its own latent representations.

Reconstruction energy

Anomalies produce high energy scores when the predictor fails to match the encoder output.

Continuous adaptation

Fine-tune the predictor head on each machine to capture unique operational patterns.

Real-time Detection

Sub-50ms inference at the 99th percentile. Canary Edge is optimized for production workloads that demand low-latency anomaly detection.

Batch mode

Submit an entire time series and receive anomaly flags for every point. Ideal for historical analysis and backfilling.

Streaming mode

Send the latest data point and get an instant determination. Designed for live monitoring and alerting pipelines.

Production-grade SLA

Engineered for 99.9% uptime with horizontal scaling and graceful degradation under load.

Inference Latency (p99)

Canary Edge

48ms

Azure AD

220ms

AWS Lookout

350ms

Open Source (PyOD)

480ms

Throughput

12,000

requests/sec (batch)

45,000

requests/sec (streaming)

Sensor Model

High-frequency

Vibration, temperature, pressure sensors sampled at sub-second intervals. Captures fast transient anomalies.

2048-pt windowFFT featuresSub-second granularity
2048point context window

Metrics Model

Low-frequency

Business metrics, daily aggregates, hourly counters. Optimized for trend shifts and seasonal deviation.

128-pt windowTrend featuresMinute+ granularity
128point context window

Multi-resolution Models

One size does not fit all time series. Canary Edge ships two model variants optimized for different data characteristics and automatically selects the right one based on your series.

Automatic selection

The API inspects granularity and sample rate to route your data to the optimal model. No configuration needed.

Tuned architectures

Each variant has a purpose-built encoder depth, context window, and feature extraction pipeline.

Unified API

Both models expose the same request/response contract. Swap between them transparently.

Per-machine Fine-tuning

Start with a generic pre-trained model, then fine-tune the predictor head on each individual machine. The system learns what "normal" looks like for your specific equipment.

Zero-shot baseline

The generic model provides accurate detection out of the box. No training period required to start.

Automatic adaptation

As data flows in, the predictor head adapts to each machine's unique patterns and operating modes.

Continual improvement

Detection accuracy improves over time as the fine-tuned model captures more operational context.

Fine-tuning Progression

1

Generic Model

82%

Pre-trained encoder and predictor provide a strong baseline for all machine types.

2

Initial Fine-tune

91%

After 24 hours of data, the predictor head begins adapting to machine-specific patterns.

3

Mature Model

97%

After 7 days, the model captures seasonal patterns, maintenance cycles, and load variations.

HEALTHY

z < 1.0

Normal operating conditions. Energy scores within one standard deviation of the running baseline.

ACTIVE

1.0 <= z < 2.0

Elevated activity. Mild deviations that may indicate changing load, warm-up, or non-critical drift.

TRANSITION

2.0 <= z < 3.0

Warning zone. Significant deviation from normal behavior. Investigate or prepare for intervention.

SHOCK

z >= 3.0

Critical anomaly. Extreme energy spike beyond three standard deviations. Immediate attention required.

Regime Classification

Beyond binary anomaly flags, Canary Edge classifies every data point into one of four operational regimes. This gives operators context-aware severity levels instead of just "anomaly or not."

Z-score thresholds

Regimes are determined by the standardized distance of the energy score from the running baseline mean.

Actionable severity

Each regime maps to a clear operational response: monitor, investigate, or act immediately.

Temporal tracking

Track regime transitions over time to identify degradation trends before they become critical failures.

How Canary Edge Compares

A side-by-side comparison of anomaly detection platforms.

FeatureCanary EdgeAzure ADAWS LookoutOpen Source
Azure API Compatible
Self-supervised (no labels)
P99 Latency
<50ms
~220ms
~350ms
varies
Per-machine Fine-tuning
manual
Regime Detection
Streaming + Batch
Multi-resolution Models
Starting Price
$0 free tier
Retired
$0.05/1K inf.
Self-host

Ready to migrate?

Switch from Azure Anomaly Detector in under 90 seconds.