Guest-facing AI that never invents a booking
A concierge bot that hallucinates a refund policy or leaks a passport number creates real liability at booking volume. EvalGuard keeps card and traveler data out of third-party LLMs, gates every answer on itinerary faithfulness, and caps runaway booking-bot spend before it compounds.
What ships today
Honest posture, not roadmap promises
Every checked item is in production today. In-progress items are flagged explicitly — no overclaiming, no vapor.
Built for buyer reality
Travel & hospitality AI use cases we ship for
Booking + reservation assistant
Guests book, change, and cancel through an AI assistant that reads fare rules and rate plans. A hallucinated waiver or invented availability is money out the door on every conversation.
EvalGuard features
- Faithfulness scorer checks answers against fare rules + rate plans
- Output guardrails block off-policy commitments — price guarantees, legal promises — in block mode
- Card-data firewall redacts PANs before any third-party model call
- Cost-cap kill switch stops runaway booking loops before they compound
In-stay concierge + service recovery
A property assistant handles requests, complaints, and comps. It must stay inside empowerment limits, de-escalate rather than inflame, and never disclose another guest's stay details.
EvalGuard features
- Toxicity scorers gate every response in block mode before it reaches the guest
- Guest-PII firewall prevents cross-guest disclosure in shared contexts
- Prompt-injection defense: 300+ attack plugins cover extraction of other guests' data
- Audit log records every comp + recovery decision for revenue review
Disruption communications at scale
Cancellations trigger thousands of personalized rebooking messages in minutes. Rights and compensation language must match regulation exactly — EU 261 mistakes at that volume become a class action.
EvalGuard features
- Guardrails block off-policy compensation promises; faithfulness scorers ground rights language in your approved policy corpus
- Faithfulness scorer verifies rebooking options against real inventory data
- Gateway cost ledger caps daily spend so IRROPS surges can't compound
- Multi-model routing spreads surge volume across providers when demand spikes
Loyalty-account assistant
Members query balances, status, and redemptions. Loyalty fraud rings probe these bots for account-takeover paths — every account-touching flow needs injection defense and rate limits.
EvalGuard features
- Per-route rate limits blunt enumeration against member-lookup flows
- Red-team packs probe redemption flows for privilege-escalation paths
- Member-PII redaction (name, contact, card data) before any third-party model call
- Consent gates honor member marketing preferences before generation
Wire it in 60 seconds
Wrap your OpenAI client. Keep card data out of the prompt.
Redaction rules + policy corpora + budget caps are configured once in the EvalGuard control plane. Your code only wraps the client.
import OpenAI from "openai";
import { wrapOpenAI, EvalGuardViolationError } from "@evalguard/openai";
const openai = wrapOpenAI(new OpenAI(), {
apiKey: process.env.EVALGUARD_API_KEY!,
projectId: "concierge-bot",
metadata: { vertical: "travel", pci: true },
blockOnViolation: true, // refuse card-data leaks
evalOnResponse: { failOnScore: 0.7 }, // policy-faithfulness gate
onViolation: (r) => alertGuestOps(r.violations),
});
try {
await openai.chat.completions.create({
model: "gpt-4o",
messages: [{ role: "user", content: guestRequest }],
});
} catch (err) {
if (err instanceof EvalGuardViolationError) {
// Block recorded in the guest audit trail. Replay via audit ID.
}
}wrapOpenAI for wrapAnthropic.Stack
Six surfaces, one platform
Eval, firewall, red-team, audit, BYOK, dashboard — every surface ships out of the box. No bolt-on vendors, no procurement cycle per capability.
Ready to ship guest AI you can defend?
Free trial includes the card-data firewall, faithfulness scorers, and the cost-cap kill switch. Point it at your fare rules and try to make it invent a waiver.
Apache-2.0 source · SOC 2 evidence engine live · full trust center