All articles
Engineering
7 min read
Guardrails that don’t slow you down
Inline policy checks for PII, jailbreaks and custom rules — without adding a second round trip.
Arjun Mehta
Safety as a stream, not a gate
Traditional guardrails wait for the full response, then run a second model over it. That doubles latency and still leaks tokens to the user while streaming. We run policy checks on the token stream itself.
Detectors are small, specialised and composable. PII redaction, prompt-injection detection and your own regex or semantic rules run side by side.
Everything is logged
Each intervention is attached to the trace with the rule that fired and the tokens it touched — exactly what your compliance team wants to see.
Next article