Guardrails & Safety
AI guardrails and safety: content filtering, output validation, safety classifiers, and responsible AI deployment.
Last updated
After this section you can
- Design the input/output classifier sandwich with a per-stage latency budget
- Compare toxicity/safety tools on latency, accuracy, and cost
- Implement a declarative policy engine and a production async safety pipeline
- Justify a build-vs-buy guardrails stack for a given industry
SD-13
Guardrails & Safety
The production safety stack: input/output classifiers, policy engines, and build-vs-buy decisions with real latency numbers.
Input & Output Classification Pipeline
A production LLM system requires safety checks at two critical points: before the prompt reaches the model (input classification) and after the model generates a response (output classification). These form a "sandwich" architecture that defends against both malicious inputs and harmful outputs.
Input/Output Safety Pipeline (Sandwich Architecture)
User Input
→
INPUT CLASSIFIERS
PII · Jailbreak · Topic · Injection
→
Pass?
→
LLM
or
Reject
LLM
→
OUTPUT CLASSIFIERS
Toxicity · Factuality · Policy · PII Leak
→
Pass?
→
Safe Response
or
Fallback