Data Security Architecture
How DULTRA protects sensitive data across every layer of the platform — from real-time voice interactions to enterprise document uploads and analysis.
Overview
DULTRA processes real-time voice conversations and enterprise document uploads. These interactions may contain personally identifiable or commercially sensitive information (PII), including national identification numbers, contact details, and financial identifiers.
Our security architecture ensures that sensitive data is identified and neutralized before reaching databases or external AI models. Corporate data, document uploads, and analysis are strictly isolated, ensuring the highest level of security.
Multi-Layer Redaction Architecture
Every piece of text or document that flows through DULTRA passes through a two-stage de-identification pipeline before persistence or forwarding to AI models.
Detected Identifier Types
Configurable Per Organization
Organization administrators can configure which identifier types are redacted. For example, sales teams that need to capture emails or phone numbers for CRM notes can selectively disable redaction for those categories. National identification numbers and financial data remain always protected and cannot be overridden.
Data Flow & Retention
Real-Time Voice Sessions
Audio streams are processed ephemerally. Voice data exists only for the duration of the active session and is never written to disk. Upon session termination, all audio buffers are destroyed with no recoverable trace.
Enterprise Document Upload & Vectorization
Corporate document uploads are processed securely and vectorized (embeddings) exclusively for your organization. This data is used only as context for your queries, ensuring your company's intellectual property is fully protected.
Transcripts & Evaluations
Text transcriptions pass through both redaction stages before database persistence. Stored transcripts contain only de-identified text with typed placeholders. Evaluations and coaching insights are generated exclusively from safe, de-identified data.
Infrastructure & Data Sovereignty
European Data Residency
- All AI inference runs on European Google Enterprise Agent Platform servers
- Database hosted within strict EU jurisdiction
- Data transfer outside the EU is architecturally prohibited
Network Security
- Advanced encryption on all communication channels
- Strict Content Security Policy (CSP)
- Automatic abuse protection and API rate limiting
AI Model Security
No Training on Customer Data
DULTRA uses Google Enterprise Agent Platform under enterprise (Enterprise) contract terms. Customer interactions and uploaded corporate documents are never used to train, fine-tune, or improve foundation models.
Ephemeral Session Context
Real-time voice sessions use stateless connections. Upon session termination, the entire conversational context is destroyed on the server side. No session resumption, no cached context, no recoverable trace.
Access Control & Authentication
Organization Isolation
Every database query is scoped to the authenticated user's and organization's identifier. Organization-level data is strictly isolated, ensuring no other company can access your information or uploaded documents.
Role-Based Access
Employees can only access their own session history and evaluations. Administrators manage team scenarios and knowledge bases exclusively within their organization boundary.
Regulatory Compliance
Data minimization and privacy by design. Only de-identified data is persisted; raw PII is never stored.
Cross-border transfer restrictions satisfied through mandatory EU regional endpoints and data residency enforcement.
Right of access and right to erasure. User data is scoped, exportable, and deletable on request.
Research
For a detailed technical analysis of the architectural principles underlying this security framework, see our published research:
Multimodal Enterprise AI in Regulated Industries: Architecting Secure, GDPR-Compliant ArchitecturesQuestions about our security?
We're happy to discuss our architecture with your security team.
Contact Security Team