# Deploy Models To Production

Ship computer vision models to serverless infrastructure. Auto-scaling, zero cold starts, and full observability out of the box.

## Uptime SLA

99.9%

## Latency P95

<100ms

## Auto-scaling

0→∞

[Deploy Your First Model](/content/demo/index.html) [Documentation](https://documentation.picsellia.com/reference/deployment)

## Infrastructure

### Serverless model serving

Deploy models without managing servers. Picsellia handles container orchestration, GPU allocation, load balancing, and auto-scaling automatically.

#### GPU & CPU Inference

Choose the right compute for your model — from T4 GPUs to cost-efficient CPU instances.

#### Container Orchestration

Automatic containerization with optimized runtimes for ONNX, TensorRT, and PyTorch.

#### Secure Endpoints

API key authentication, rate limiting, and encrypted traffic by default.

### Deployment Architecture

- Managed infrastructure
- API Gateway
- Load balancer + Auth
- HTTPS
- API Keys
- Rate Limiting
- Inference Servers
- Auto-scaled replicas

#### Model Registry

- Versioned artifacts
- Monitoring
- Predictions logged

## Developer Experience

### Deploy in a few lines of code

Use the Python SDK to deploy, update, and manage models programmatically. Full API access for CI/CD integration.

#### Deploy A Model

```python
# Connect and get deployment
from picsellia import Client

client = Client()

# Create deployment with model
deployment = client.create_deployment(
  name="prod-v3"
)
deployment.set_model(model_version)
```

#### Run Inference

```python
# Run prediction from file path
result = deployment.predict(
  "image.jpg"
)

# Run prediction from bytes
result = deployment.predict_bytes(
  "image.jpg",
  raw_image
)

# Send to monitoring
deployment.monitor("image.jpg")
```

### REST API

```bash
# Direct API call
curl -X POST "https://serving.picsellia.com/v1/predict" \
  -H "Authorization: Bearer $API_KEY" \
  -F "image=@photo.jpg" \
  -F "deployment_id=dep_abc123"
```

## Auto-Scaling

Active

### Replica count over 24h

| Time  | Replicas |
|-------|----------|
| 06:00 | 1        |
| 09:00 | 3        |
| 12:00 | 6        |
| 15:00 | 4        |
| 18:00 | 2        |
| 22:00 | 1        |

- Min: 1 replica
- Max: 6 replicas
- Cost optimized

### Scale to match demand

Automatically scale from zero to thousands of requests per second. Pay only for the compute you use, with intelligent scaling policies.

- Scale-to-zero for cost efficiency
- CPU and request-based scaling policies
- Configurable min/max replica bounds
- Cold-start optimization with warm pools
- Multi-region deployment support

## Built for Production

### Everything you need to serve models

From model registry to production endpoint, Picsellia handles the entire deployment lifecycle with enterprise-grade reliability.

#### Model Registry Integration

- Deploy any model version from your registry.
- Full lineage from experiment to production endpoint.
- Version management
- Artifact tracking
- Rollback support

#### Runtime Optimization

- Automatic model optimization with ONNX Runtime, TensorRT, or custom serving containers.
- ONNX Runtime
- TensorRT acceleration
- Custom containers

#### Monitoring Built-In

Every prediction is logged. Track latency, throughput, and anomalies from day one.
- Real-time dashboards
- Anomaly detection
- Drift tracking

## Ready to deploy your models?

Go from trained model to production endpoint in minutes. Serverless, scalable, and fully managed.

[Start Free Trial](/content/trial/index.html) [Request Demo](/content/demo/index.html)