Defend Overview
Safeguard your AI applications in production using DeepRails Defend—powered by real-time guardrails and intelligent quality filters.
DeepRails Defend is your real-time protection layer for generative AI in production. It automatically detects and filters low-quality and unsafe outputs using Guardrail metrics and intelligent thresholds. Whether you’re deploying an LLM in a user-facing product or mission-critical backend system, Defend ensures only high-quality, safe generations reach your users.
DeepRails Defend can help safeguard your users and systems from a wide range of quality and safety risks:
Core Features
-
Guardrail-Based Filtering
Intercept outputs using trusted DeepRails Guardrail metrics like correctness completeness, adherence and safety. -
Intelligent Thresholds
Choose between automatic defense (AI-optimized thresholds) or manually configure custom filters for your use case. -
Real-Time Enforcement
Stop unsafe or low-quality completions before they reach users, integrated seamlessly into your production stack. -
Insights & Analytics
Explore triggered defenses, audit decisions, and refine thresholds using Krino Console’s Defend insights.
Krino Console: Defend Dashboard
The Defend Workflow
Configure Your Defend Workflow
Create a Defend Workflow in the Krino Console by selecting Guardrail metrics you want to enforce. Choose between Automatic Defense, which intelligently sets thresholds, or define your own Custom Thresholds tailored to your content and risk profile.
Iterate and Test Your Conditions
Before going live, test your Defend setup against known evaluation sets. Validate that your workflow is neither over-triggering nor too lenient. Fine-tune thresholds and metric selection until you’re confident in your coverage.
Log Production Traffic
Once tuned, begin logging real completions from your application via API. Defend will score and filter each output in real time, ensuring only high-quality, compliant responses reach users.
Monitor and Improve
Use the Krino Console to analyze filtered outputs, review trigger patterns, and continuously refine your defense strategy as your application evolves.