LLM guardrails that hold up on your fine-tune.

Distilled classifier heads: 5-15ms per call on a T4 or L40S. Llama Guard 3 (8B) inline: 50-150ms depending on context length. For latency-sensitive routes we run a small classifier inline and a larger one async on a sample. Output guardrails run streaming - they stop generation on the offending token rather than waiting for the whole response.

Two ways. First, our red-team runs domain-specific attacks during the build - we don't ship without exploring the prompt space your customers actually inhabit. Second, the production system logs flagged inputs under a strict review policy, our team triages new attack patterns weekly, and the input classifier gets retrained on the new corpus.

Yes - and the threat model is meaningfully different. The dominant attack against production agents isn't a jailbreak in the user message; it's an instruction injected through a retrieved document, an MCP server response, or a tool's return value. We add input scrubbing at the agent tool boundary, structured-output validation on every tool call, and per-tool authorization policies that the agent can't override at runtime. Agent guardrails are a superset of chat guardrails - the underlying classifiers are the same, but the instrumentation points and the red-team playbook are distinct.

Article 15 of the EU AI Act requires high-risk AI systems to include appropriate accuracy, robustness, and cybersecurity measures - with technical documentation that supports post-market monitoring. A calibrated classifier suite with per-attack-class F1 measurements, a structured red-team report, and a CI-gated eval suite are the evidence artifacts that satisfy both internal audit and third-party conformity assessments. We produce these as part of every engagement; the documentation format is designed to map to Article 9 risk management and Article 15 technical requirements, not to serve as a marketing brochure.