Live Monitoring Active

Your AI asked for a password?
Sentinel blocked it.

9-layer semantic pipeline that detects dangerous AI responses before they reach real customers. 100% recall on policy violations. 12 seconds for 5,000 tickets.

⚡ Try Live Demo
sentinel — evaluation pipeline
ticket "Customer cannot access account..."
answer "Please send your password and CVV..."
running 9-layer evaluation pipeline
layer 1 ⚠ policy match: secret_collection (0.91)
layer 3 ✗ contradiction: "do NOT request" vs "request"
layer 5 score: 0.04 risk: 0.96
⊘ REJECTED — blocked before customer sees it
Secret collection: answer asks for password and CVV
Critical contradiction detected with safety policy
0% policy coverage — missed all required directives
100%
Recall on violations
12s
5,000 tickets on CPU
82%
Overall accuracy
HOW IT WORKS
Three seconds to a verdict
01

Paste any AI answer

Drop in the ticket, the model's response, and the expected policy. No setup needed. Works offline with the local evaluator.

02

9-layer pipeline runs

Semantic matching, policy coverage, contradiction detection, severity scoring — all in under 10ms. No black box.

03

Clear verdict + reason

Release, Manual Review, or Rejected — with exactly why, in plain English. Every decision is auditable and exportable.

THE PROBLEM IT SOLVES
What happens without Sentinel

✕ Without Sentinel

→Dangerous answer reaches the customer
→CVV / password requests slip through keyword filters
→Model drift invisible until incident happens
→No audit trail, no rollback, no evidence bundle

✓ With Sentinel

→Answer blocked in 5ms before customer sees it
→Semantic pipeline catches meaning, not just keywords
→Drift monitored continuously per category
→Full audit bundle, operator decisions, export JSON
THE ENGINE
9-layer semantic evaluation
LAYER 1
Semantic
Similarity
LAYER 2
Policy
Coverage
LAYER 3
Credential
Detection
LAYER 4
PII
Detection
LAYER 5
Hallucination
Detection
LAYER 6
Contradiction
Detection
LAYER 7
Adversarial
Guard
LAYER 8
Risk
Aggregator
LAYER 9
Confidence
Engine
Live Sandbox — Score any answer now ● pipeline_status: active
Customer Ticket
Model Answer to Check
Expected Answer (Gold Directive)
Policy Excerpt (Constraints Context)
🛡️
Active Pipeline Verdict

Enter a candidate support answer above and click score to run the SentenceTransformers evaluation gate.

BENCHMARKS
Sentinel vs alternatives
Approach Recall (dangerous) Overall Accuracy False Positive Rate Avg Latency
🛡 LLM Sentinel Pro 100% 82% 40% 10.59ms
Keyword Filter ~60% ~70% 12% 0.1ms
No Monitoring 0% — 0% 0ms
CONSOLE PREVIEW
Simplified. Focused. Fast.
Score Ticket
Dashboard
Datasets
Advanced ↓
Settings
Customer Ticket
Model Answer to Check
Account Access Advanced: Context
pipeline_status: active
REJECTED
0.04
SCORE
0.96
RISK
4
CLAIMS
Block
DECISION
Secret collection: password + CVV
Contradiction with safety policy
0% policy directive coverage

Run it in 3 commands

Open-source. No account needed. Runs fully offline on your machine.

# Install dependencies
pip install -r requirements.txt
# Start the server
python -B backend/server.py
# Open in browser
open http://127.0.0.1:8000
View on GitHub