AIGRaaS
Product

Three Evaluation Modes

System prompt, pre-delivery, or post-delivery — use one or all three

Mode 1 injects constitutional rules into your AI's system prompt (0ms). Mode 2 evaluates every response before delivery (<10ms). Mode 3 audits asynchronously. Configure once, deploy everywhere.

0ms

Mode 1 latency

<10ms

Mode 2 latency

Async

Mode 3 latency

Try in PlaygroundHow it Works

Different apps need different latency budgets. A voice AI can't afford a 10ms round-trip on the hot path; a nightly compliance review doesn't care about latency at all. One ruleset, three deployment modes — use whichever fits, stack them for defense in depth.

Configure once, deploy in layers

Your ruleset is the source of truth. Modes decide how it's enforced: injected into the system prompt for zero-cost prevention, evaluated before delivery for real-time blocking, or audited after the fact for observability. Use all three together for maximum coverage.

mode-config.ts
import { AIGRaaS } from "@aigraas/client";

const guard = new AIGRaaS({ apiKey: process.env.AIGRAAS_KEY });

// Mode 1 — generate system prompt at startup
const systemPrompt = await guard.generateSystemPrompt({
  ruleset: "healthcare-intake-v1",
});

// Mode 2 — pre-delivery evaluation in the hot path
const verdict = await guard.evaluate({
  ruleset: "healthcare-intake-v1",
  response: aiResponse,
  mode: "pre-delivery",
});

// Mode 3 — post-delivery audit, fire and forget
guard.audit({
  ruleset: "healthcare-intake-v1",
  response: aiResponse,
  delivered: true,
});

Mode 1 — System Prompt Injection

Zero-latency prevention. AIGRaaS generates constitutional rules your AI provider enforces as part of its normal operation. No API calls needed at runtime.

Mode 2 — Pre-Delivery Evaluation

Under 10ms. Every AI response passes through the evaluation engine before reaching the user. Blocks, redirects, or approves in real time.

Mode 3 — Post-Delivery Audit

Async. Responses ship immediately and the audit runs in the background. Violations show up in the dashboard and can trigger alerts.

Defense in depth

Stack all three. Mode 1 prevents, Mode 2 blocks, Mode 3 catches anything that slipped through. One ruleset, three checkpoints.

The three modes are not alternatives — they're layers. Use Mode 2 on the hot path, Mode 3 for forensics, Mode 1 as a cheap first line.

“Mode 2 on the hot path plus Mode 3 as a safety net means we get both real-time blocking and a queryable audit trail. We stopped tuning the system prompt six months ago — the ruleset does that job.”

— Production architect, Voice AI operator

Questions we get

Ready to try it?

Evaluate any AI response in the playground — no signup required.

Try in PlaygroundView Pricing