AI safety + compliance for the grid
AI that reads outage tickets, SCADA summaries, or switching procedures sits next to critical infrastructure. EvalGuard keeps grid data inside your boundary with air-gapped self-hosting, gates operational answers on runbook faithfulness, and logs every interaction for CIP-style audit review.
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
Energy & utilities AI use cases we ship for
Switching-procedure and runbook copilot
Field crews and control-room operators query an assistant grounded in switching orders and O&M runbooks. A hallucinated isolation step is a safety event — every answer must be faithful to the current procedure revision.
EvalGuard features
- Faithfulness scorer checks each step against the source procedure
- Hallucination gate blocks answers that drift from the runbook
- Audit log ties every answer to procedure revision for post-event review
- Self-hosted deploy keeps switching data inside the ESP boundary
Outage-communications drafting
AI drafts customer notifications and regulator updates during outages. Estimates must match the OMS record — an invented restoration time becomes a regulatory complaint at scale.
EvalGuard features
- Faithfulness scorer checks restoration estimates against the OMS record
- Toxicity scorers + brand-voice guardrails keep mass notifications on-script
- Output guardrails in block mode reject regulator updates that introduce un-cited claims
- Gateway cost ledger caps spend during high-volume outage events
Customer-service assistant with account access
The support bot reads billing, usage, and payment-plan records. Usage data reveals occupancy patterns — it must be redacted from third-party model logs and protected from cross-account extraction.
EvalGuard features
- Data firewall redacts account numbers, addresses, and service identifiers
- Prompt-injection defense: 300+ attack plugins cover cross-account extraction
- Per-route rate limits blunt enumeration against account-lookup flows
- Consent gates honor customer data-sharing preferences before generation
Regulatory-filing summarization
Rate-case teams summarize dockets, testimony, and precedent into filing drafts. Citations must point at real docket entries — a fabricated citation in a FERC filing is a credibility disaster.
EvalGuard features
- Citation-format scorer verifies every reference against the source corpus
- Faithfulness gate rejects summaries introducing un-cited claims
- Per-project BYOK isolation separates rate-case workspaces
- Evidence-bundle export documents the AI-assist trail for counsel
Wire it in 60 seconds
Wrap your OpenAI client. Keep grid data in the perimeter.
Redaction rules + faithfulness 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: "ops-runbook-copilot",
metadata: { vertical: "utilities", nercCip: true },
blockOnViolation: true, // refuse asset-data leaks
evalOnResponse: { failOnScore: 0.8 }, // procedure-faithfulness gate
onViolation: (r) => alertGridSecOps(r.violations),
});
try {
await openai.chat.completions.create({
model: "gpt-4o",
messages: [{ role: "user", content: switchingQuery }],
});
} catch (err) {
if (err instanceof EvalGuardViolationError) {
// Block recorded in the 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 grid-adjacent AI you can defend?
Free trial includes the data firewall, faithfulness scorers, and the tamper-evident audit trail. Air-gapped self-host available for OT-adjacent workloads.
Apache-2.0 source · SOC 2 evidence engine live · full trust center