Deployment Modes
Blocking, monitoring, async, and self-hosted — pick your risk posture
Deployment Modes
SentinelAI adapts to your risk posture and throughput.
The four modes
| Mode | Use case | Behavior |
|---|---|---|
| Blocking | Customer-facing chat, copilots | Verify every response before it hits your user; serve, correct, or block |
| Monitoring | Evaluation, staging, low-risk features | Log everything, review flagged ones later |
| Async | High-throughput pipelines | Fire-and-forget analysis with webhook callbacks |
| Self-hosted | Regulated industries, data residency | Everything runs on your own infrastructure |
Blocking
The default. Your app sends the prompt/response pair, SentinelAI returns the verdict, and your app acts on it — all before the user sees anything.
result = client.verify(prompt=prompt, response=llm_output)
if result["status"] == "hallucinated":
return result["corrected"] # serve the fix, not the flawMonitoring
Log everything, act later. Perfect for post-hoc analysis, drift detection, and compliance auditing without adding latency to the request path.
Async
For high-throughput workloads where you can't wait for a verdict inline:
- Fire-and-forget the analysis
- Receive the result via webhook
- Store and aggregate at your own pace
Self-hosted
The open-source core runs entirely on your infrastructure. Data never leaves your network. The same SDK, the same detectors, the same API — just your hardware.
Self-hosting is free and open source. The managed cloud adds dashboards, alerting, and settings management on top of the same core.