In-vehicle AI with a safety case behind it
An in-cabin assistant that misreads an owner's-manual procedure or coaches a driver through a maneuver becomes a safety finding with a VIN attached. EvalGuard gates vehicle AI on manual-faithfulness evals in CI, guards the in-cabin surface at runtime, and keeps the tamper-evident trail UNECE-style audits expect.
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
Automotive AI use cases we ship for
In-cabin voice assistant
Drivers ask about warning lights, tire pressure, and maintenance while driving. A wrong answer about a brake warning is a safety event — every response must be faithful to the owner's manual for that exact model-year.
EvalGuard features
- Faithfulness scorer checks answers against the model-year manual corpus
- Realtime voice control plane with per-frame guardrail scanning
- Hallucination gate blocks procedures that drift from the source manual
- Audit log ties every safety-relevant answer to manual version + VIN context
Release gating for assistant updates
Every OTA update to the assistant model or prompt stack risks regressions on safety-critical intents. Releases need eval gates the safety team can point to in the ISO 26262-style safety case.
EvalGuard features
- CI eval gates: safety-intent test suites block releases below threshold
- Dataset versioning keeps the safety-intent suite reproducible per release
- Red-team regression: 300+ attack plugins re-run against every candidate build
- Evidence-bundle export documents gate results for the safety case
Dealer + service copilot
Service advisors query TSBs, recalls, and repair procedures. A hallucinated torque spec or skipped recall step creates warranty and liability exposure across the dealer network.
EvalGuard features
- Citation scorer requires TSB/recall grounding for every procedure answer
- Faithfulness gate rejects answers not grounded in the selected model-year corpus
- Per-project BYOK isolation separates OEM + dealer-group workspaces
- Gateway cost ledger attributes spend per rooftop with budget caps
Connected-vehicle data assistant
Fleet and owner apps expose AI over telemetry — location history, driving behavior, diagnostics. That data is regulated PII in most markets and must never leak into third-party model logs.
EvalGuard features
- PII firewall redacts VIN and 18+ identifier types before LLM calls
- Consent gates honor owner data-sharing preferences before generation
- Per-route rate limits contain scraping against fleet-lookup endpoints
- Self-hosted deploy available inside the vehicle-data boundary
Wire it in 60 seconds
Wrap your OpenAI client. Gate it on the manual.
Manual corpora + safety thresholds + audit retention 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: "in-cabin-assistant",
metadata: { vertical: "automotive", modelYear: "2026" },
blockOnViolation: true, // refuse driver-PII leaks
evalOnResponse: { failOnScore: 0.8 }, // manual-faithfulness gate
onViolation: (r) => alertSafetyTeam(r.violations),
});
try {
await openai.chat.completions.create({
model: "gpt-4o",
messages: [{ role: "user", content: driverQuery }],
});
} catch (err) {
if (err instanceof EvalGuardViolationError) {
// Block recorded in the incident 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 vehicle AI you can defend?
Free trial includes the eval gates, manual-faithfulness scorers, and the incident audit trail. Self-hosted deploy available inside your vehicle-data boundary.
Apache-2.0 source · SOC 2 evidence engine live · full trust center