Most security tools read your policies at runtime. We compile them into mathematics first. The scanner never reads English — it multiplies vectors.
System Architecture
Each component does one thing and does it precisely.
Feed any policy in plain English. The Compiler embeds it using a fine-tuned model we run inside your deployment, then trains a lightweight linear classifier. Output: a 7KB binary artifact.
Numbers from production builds. Not marketing estimates.
| Metric | Value | Notes |
|---|---|---|
| Scan latency p50 | 12ms | 500-antibody library |
| Scan latency p99 | 94ms | 500-antibody library |
| Antibody size | ~7KB | float32, compressed |
| Compile time | 2.1s | single-concept policy |
| AUC (held-out) | 0.995 | insider-trading detector |
| Cold-start AUC (federated) | 0.86 | vs 0.74 without network |
The scanner is a local library call. No HTTP round-trip to our infrastructure.
# Create a scanner from your antibody library
import compiled
# Connect to your tenant (scanner runs locally)
scanner = compiled.Scanner(tenant="acme-corp")
# Score any text in under 100ms
result = scanner.scan("Forward this to my personal email before the announcement")
# → ScanResult(score=0.94, verdict="FLAGGED", antibodies_triggered=["insider-trading-v3"])
The antibody is a math artifact — 7KB traveling inward. Your messages, embeddings, and verdicts stay in your environment forever.
Into your environment
7KB antibody artifact only
Never outbound
Messages, embeddings, verdicts
Federated (opt-in)
DP centroids only, ε=1.0
The questions your security architects will ask.
We'll walk through how the Compiler, Antibody, and Scanner map to your specific threat models.
Request a technical demo