Compliance, engineered.
Named-entity redaction at ingest in 800 to 1200 ms. Real-time compliance alerts during the live call in 380 ms. Vertical-specialist MoE Operators trained in a 4,000-agent live contact center. Reference documentation for compliance officers, CISOs, and risk reviewers.
What QEval® redacts and how customers configure it.
QEval® uses Named Entity Recognition trained on millions of conversations to identify sensitive entities in three categories: PCI, PHI, and PII. Configuration is per program. Five PCI entities are required and locked. PII entities are recommended by default. Optional PII and PHI entities are opt-in. Customers choose what to mask before deployment.
PCI entities. Always on.
PII entities. On by default.
Custom PII catalog.
Custom PHI catalog.
Four stages. Watch a transcript redact in place.
Live, auto-advancing. The sample transcript redacts as the stages advance. Original media is deleted at stage three. Only the redacted transcript reaches the MoE.
Voice or chat arrives. Stored temporarily on an IP-restricted Etech server.
Interaction streams in over OAuth 2.0 from any connected CCaaS. Audio is encrypted at rest (AES-256) and held only long enough to ingest. No model has read the data at this point.
NER tags PCI, PHI, and PII entities at ingest, in 800 to 1200 ms.
Named Entity Recognition tags every entity selected in the customer's configuration. Non-sensitive numbers (ordinals, percentages, prices, times) are preserved so analytics still work. The flagged version is staged; unredacted data remains walled off from any model.
Redaction applied. Original media deleted upon successful ingestion.
Sensitive tokens become redaction markers in transcripts; numbers are masked with symbols in text and replaced with silence in audio. Per the contractual data flow, media is deleted upon successful ingestion before the next interaction is processed. Only the redacted version is retained.
Redacted transcript reaches the MoE. Every classification traces to an expert.
The proprietary closed-source Mixture-of-Experts routes each scorecard item to its specialized expert. No third-party foundation model is in the loop, for training or inference. Every classification ships with the expert that scored it, the transcript span, the confidence score, and the timestamp.
Where redaction happens. Four surfaces.
Voice audio
Sensitive numbers are replaced with silence. Non-sensitive numbers (ordinals, prices, durations) are preserved so analytics still work.
Text transcripts
Detected entities are replaced with tokenized redaction markers in text streams. Chat, email, and SMS handled the same way.
On-screen / desktop
Screen capture redaction prevents sensitive data from being recoverable through screenshots of the agent desktop during live calls.
In-flight, real-time
Live detection runs alongside post-call cleaning. Both archived audio and active interactions are covered by the same NER pipeline.
Real-time alerts while the call is still happening.
When the compliance MoE expert detects a critical violation mid-call, the alert dispatches in 380 ms to the channels each customer configures. Choose any combination: email, SMS, Microsoft Teams, Slack, or outbound voice call to a supervisor.
Critical compliance triggers
Configurable per program. Each trigger maps to one or more channels.
One MoE per vertical. One documented violation taxonomy.
Compliance scoring is not generalist work. The MoE routes each scorecard item to the expert sub-model trained on that vertical's regulations. Twelve purpose-trained compliance experts, one violation taxonomy across all of them.
Twelve vertical compliance experts.
Each expert sub-model is tuned to a vertical's named regulations. Disclosure language, advice unauthorized for the agent's role, PII mishandling, identity-verification gaps. Calibration variance under 2% against the live operation's confirmed violation register.
Eight documented violation codes.
Every flagged event maps to a named category. No miscellaneous bucket.
disclosure_missRequired statement not deliveredpii_mishandlingPII captured or repeatedunauthorized_adviceOut-of-role guidanceharmful_outputAggressive or threateningrecording_consentNotification missedidentity_verificationKYC step missedescalation_failureTrigger not actionedretention_violationStorage or DSR issueCompliance specialists trained on the regulations your agents actually face.
QEval® is operated by ETS Labs, the engineering arm of Etech Global Services, which runs a 4,000-agent live contact center across financial services, healthcare, collections, insurance, and telecom. Compliance reviewers come from live operating programs, not consulting engagements. They review violations alongside the automated scoring as an ongoing second layer.
Software companies study the regulations. Operators live with them.
Among the QA software vendors in the category, ETS Labs is the only one that simultaneously operates a live regulated contact center at scale. The compliance team's expertise is from actually handling Mini-Miranda misses, HIPAA edge cases, CMS scripted disclosures, and TCPA exposure on real calls.
Collections
Healthcare
Financial services
Insurance
Telecom
Retail / outbound
Energy + Utilities
Automotive
Government
BPO / outsourcing
Travel + Hospitality
Education
How data moves through the platform.
End-to-end data flow from source to outputs. Authentication is OAuth 2.0. Temporary audio storage is on an IP-restricted Etech-owned server, encrypted at AES-256. Media is deleted upon successful ingestion. Only the redacted version reaches the MoE.
Contact Center Platform
Voice, chat, email, SMS. OAuth 2.0 token via POST /auth/token.
Etech Server
IP-restricted. AES-256 at rest. No public endpoints.
NER Redaction
800-1200 ms. PCI / PHI / PII tagged. Original media deleted on ingest.
MoE Speech
STT, classifications, summarization, intent. Closed-source. No third-party LLM.
Scorecards + Audit
JSON to Q&A row data. Scorecards, dashboards, audit trail.
The compliance posture in one page. The audit pack in the Trust Center.
Eight named certifications, each scope-stated. The Trust Center hosts the full audit documentation under NDA: SOC 2 reports, pen-test results, sub-processor lists, BAA templates, redaction methodology, encryption attestations. Everything a CISO needs to close a vendor risk review without a follow-up call.
SOC 2 reports, pen tests, sub-processor list, audit documentation.
Hosted at trust.etslabs.ai under NDA. Everything a CISO or compliance officer needs to close a vendor risk review.
Compliance scoring is Layer 1. The audit-trail discipline scales to Layer 6.
The redaction architecture and audit-completeness that make Layer 1 defensible are the same architecture that makes Layer 6 (Strategic Intelligence) auditable across a human and AI workforce.
What a compliance officer needs to know.
What happens to the original recording after redaction?
Per the contractual data flow, media is deleted upon successful ingestion before the next interaction is processed. Only the redacted transcript is retained by the MoE pipeline and the audit log. There is no recoverable source file kept after the redaction stage completes.
Does customer data reach OpenAI, Anthropic, Google, or AWS Bedrock?
No. QEval® runs on a proprietary closed-source Mixture-of-Experts model operated by ETS Labs. No third-party foundation model is in the data path at any stage, for training or for inference. No third-party foundation model is in the data path at any stage — for training or inference — so customer data never enters an external training loop.
How does custom redaction configuration work?
Five PCI entities are required and locked. Four PII entities are recommended (on by default; opt-out per entity). Thirty-three optional PII entities and seven optional PHI entities are off by default and the customer opts in. The selections are made before deployment and confirmed with the implementation project manager.
What does the 380 ms alert latency cover?
It measures the time from the compliance MoE expert flagging a critical violation during a live interaction to the first configured channel receiving the alert. Configured channels include email, SMS, Microsoft Teams, Slack, and outbound voice call to a supervisor. Customers select any combination, and trigger rules are configurable per program.
What does the vertical compliance specialist layer actually do?
ETS Labs operates a 4,000-agent contact center across financial services, healthcare, collections, insurance, and telecom. Compliance specialists drawn from those live operating programs review flagged events and tune scoring rules against the regulations agents actually face on real calls. This runs alongside the automated scoring as an ongoing review layer, not a one-time consulting engagement.
Is automated redaction guaranteed to be 100% accurate?
No automated redaction engine is. Accuracy depends on transcription quality, which depends on audio quality, accent, speech rate, and background noise. QEval® ships transcription optimization, calibration during deployment, and a human-in-the-loop review posture as the answer. A third-party audit at scale found no exploitable information was retained after the redaction process. Audit documentation is available in the Trust Center under NDA.
Is QEval® ready for the August 2, 2026 EU AI Act deadline?
Yes. ISO 42001 (AI management system) is held; QEval® is among the first QA-category platforms to surface it. Customer-facing emotion recognition is not used, keeping QEval® clear of the EU AI Act's prohibited-AI tier. The Trust Center hosts the conformance documentation under NDA.
Score a live call against your compliance.
Bring your scorecards, your compliance rules, your vertical regulations. QEval® scores a live call against them, ships audit documentation to your CISO, and pilots toward ROI in 120 days, contractually.