# Supervisez vos modèles  en production

Observabilité en temps réel pour les modèles de vision par ordinateur. Détectez les anomalies, suivez le drift et fermez la boucle de feedback automatiquement.

### 24/7 Monitoring

<1% Anomaly Rate  
<50ms Latency P99  
[See It In Action](/content/demo/index.html) [Documentation](https://documentation.picsellia.com/docs/deployment)

## Anomaly Detection

### Catch failures before users do

Automatic 24/7 monitoring flags anomalous inputs and low-confidence predictions. Reveal blindspots and discover edge cases in production data.

#### Low Confidence Alerts

Flag predictions below confidence thresholds automatically

#### Distribution Shift

Detect when production data drifts from training data

#### Novel Patterns

Discover previously unseen data patterns and edge cases

### FLAGGED PREDICTIONS Last hour  
  
23%  
  
45%  
  
31%  
  
89%  
  
18%  
  
52%

## Live Dashboard

### Real-time prediction insights

Filter millions of inferences to identify the top anomalies. From edge case detection to training dataset integration in seconds.

1.2M  
### Total Inferences Last 24 hours

47ms  
### Avg Latency P50 response time

99.1%  
### Accuracy Based on reviewed predictions

## MONITORING METHODS

```python
# Monitor from file
deployment.monitor(
  "image.jpg",
  tags=["production"]
)

# Monitor from bytes
deployment.monitor_bytes(
  "frame.jpg",
  raw_bytes
)
```

```python
# Get deployment stats
stats = deployment.get_stats(
  window="24h"
)

# Access metrics
print(stats.predictions)
print(stats.reviews)
print(stats.latency_p99)
```

## DRIFT DETECTION vs. training baseline

- Mean Brightness: 127.3 ↑7.4%
- Contrast: 0.4 ↓6.7%
- Class Distribution: 23.1% ↓10.5%
- Object Size (avg): 156.0px ↑9.9%
- Confidence (avg): 0.8 ↓7.7%

3 metrics drifting

### Data Drift

## Track distribution changes

Compare production data against your training baseline. Get alerted when distributions shift beyond acceptable thresholds.

- Image statistics (brightness, contrast, blur)
- Class distribution changes
- Confidence score degradation
- Custom threshold alerts

## Continuous Improvement

### Close the feedback loop

Convert production failures into training data. Review predictions, attach to datasets, and trigger retraining automatically.

- Monitor Production inference
- Detect Anomalies & drift
- Review Verify & label
- Retrain Improve model

### Continuous Training

Automatically trigger retraining when prediction review thresholds are met. Keep models fresh with production data.

`deployment.toggle_continuous_training()`

### Continuous Deployment

Manage model promotion policies between staging and production. Shadow model support for A/B testing.

`deployment.set_shadow_model(new_version)`

## Ready to monitor your models?

Start detecting anomalies, tracking drift, and improving models with production feedback.