CONTEXT
Missing organizational knowledge
Generic models do not automatically understand a company’s terminology, approved translations, product names, or domain conventions.
CDTI industrial research and development
Privacy-Controlled AI Translation: the research behind adaptive, measurable, and publishable multilingual workflows
Pangeanic’s CDTI project combined controlled information retrieval, adaptive machine translation, automatic quality estimation, and AI-assisted review to solve a production problem: how to translate enterprise content securely, apply organizational knowledge consistently, and determine what can be published automatically.
IDI-20240301 CDTI project identifier
18 months January 2024 to June 2025
€476,487 Total published investment
€184,237 Published ERDF contribution
CDTI and ERDF Spanish R&D with European co-financing
The production problem
Fluent translation is insufficient when organizations also require terminology, privacy, consistency, and release control
General-purpose translation systems can produce convincing language, but enterprise publication depends on more than fluency. Technical terminology must remain stable, approved expressions must be respected, sensitive information must stay within controlled environments, and quality decisions must be repeatable.
Conventional workflows often separate translation from terminology, reference materials, quality assurance, and publishing. This creates manual handoffs, inconsistent review decisions, and unnecessary human effort.
The CDTI project addressed these limitations by researching a connected architecture in which translation can retrieve approved organizational knowledge, adapt to domain requirements, estimate its own quality, and trigger the appropriate next action.
PRIVACY
Sensitive content exposure
Legal, technical, financial, medical, and public-sector documents require controlled processing and clear data boundaries.
QUALITY
Uniform review is inefficient
Reviewing every sentence wastes expert time, while publishing everything automatically introduces unacceptable risk.
RELEASE
Publishing lacks objective gates
Production systems need measurable rules for automatic release, targeted review, corrective processing, or rejection.
Research architecture
Retrieval, adaptation, quality estimation, and corrective processing inside one translation workflow
The project investigated how approved enterprise information could guide translation while automated quality signals controlled the transition from generation to publication.
01
Retrieve approved knowledge
Search translation memories, terminology, bilingual references, style information, and domain resources relevant to the source content.
02
Generate an adapted translation
Apply retrieved context to produce output aligned with organizational terminology, tone, subject matter, and previous linguistic decisions.
03
Estimate translation quality
Score output without requiring a human reference translation and identify segments that may need correction or review.
04
Route the next action
Publish, post-edit, send for expert review, or block the output according to quality, risk, content type, and workflow policy.
From research project to product
The CDTI project became a production architecture for adaptive translation and intelligent multilingual publishing
The funded research did not remain a standalone prototype. Its results strengthened the technology behind Deep Adaptive AI Translation and its integration into the ECO Intelligence Platform.
Deep Adaptive AI Translation provides the generation layer. It retrieves and applies approved terminology, translation memories, domain knowledge, style instructions, and previous linguistic decisions while the translation is being produced.
MTQE provides the measurement and control layer. ECO provides the operating layer that receives documents, applies privacy controls, preserves formats, manages APIs, routes content, and determines whether translated material should be published, corrected, or reviewed.
RESEARCH
CDTI project
Investigated retrieval, adaptation, quality estimation, corrective processing, privacy, and workflow control.
TRANSLATE
Deep Adaptive AI Translation
Generates enterprise translations using approved linguistic assets and dynamically retrieved organizational knowledge.
EVALUATE
MTQE
Scores translation quality and provides the evidence required for automatic release or selective review.
OPERATE
ECO Intelligence Platform
Orchestrates documents, privacy, translation, quality gates, corrective actions, APIs, and multilingual publication.
The translation product layer
Deep Adaptive AI Translation turns the project’s research into a deployable enterprise translation product
Deep Adaptive AI Translation is the production expression of Pangeanic’s work on contextual adaptation. Instead of treating translation as a single generic model call, it incorporates the linguistic assets and preferences that define how an organization communicates.
Translation memories, terminology, approved sentence pairs, CSV resources, style instructions, and domain knowledge can influence translation dynamically. This reduces dependence on static generic behavior and improves consistency across recurring enterprise content.
The result is a translation layer designed for organizations that require repeatability, traceable adaptation, terminology control, and integration with document, localization, or content-management workflows.
MEMORY
Approved bilingual assets
Reuse reviewed translation memories and sentence pairs as active evidence for new translation decisions.
TERMS
Terminology enforcement
Apply product names, domain terminology, acronyms, protected terms, and organization-specific equivalents consistently.
STYLE
Tone and communication rules
Align output with approved style, audience expectations, domain conventions, and previous human choices.
ADAPT
Dynamic contextual adaptation
Retrieve the information required for each document, paragraph, or segment rather than relying only on a static model profile.
MTQE as a control layer
Quality estimation turns translation confidence into an operational publishing decision
MTQE estimates translation quality without waiting for a human reference. In production, the score becomes useful when it controls what happens next.
HIGH
Publish automatically
High-confidence content can continue to publication or downstream processing when it satisfies the organization’s quality and risk policies.
MEDIUM
Apply corrective processing
Borderline segments can pass through automatic post-editing, alternative generation, terminology verification, or additional model checks.
REVIEW
Route selectively to experts
Human reviewers receive the content where their expertise adds the most value instead of reviewing every translated sentence.
LOW
Block unsafe publication
Low-confidence or high-risk output can be stopped automatically before it reaches a public website, customer, employee, or regulated workflow.
TRACE
Record quality evidence
Scores, routing decisions, corrections, and human feedback provide evidence for auditing and continuous workflow improvement.
LEARN
Improve future decisions
Reviewed outputs and error patterns can strengthen evaluation sets, terminology resources, adaptation logic, and model-selection policies.
The operational layer
ECO turns research components into a usable enterprise translation and document workflow
A model, retrieval component, or quality score has limited value in isolation. Organizations require an environment where users can upload documents, select services, apply privacy controls, manage terminology, preserve formats, call APIs, and route content according to quality.
ECO Intelligence Platform provides that operating layer. It connects translation, document processing, anonymization, multilingual retrieval, MTQE, APIs, and governed execution within controlled environments.
The relationship is direct: Deep Adaptive AI Translation generates context-aware output, MTQE evaluates operational risk, and ECO executes the document and publishing workflow.
Adaptive translation Apply terminology, approved language, domain context, and organizational preferences MTQE quality gates Estimate quality and control automatic publication, correction, or review Privacy controls Mask or anonymize sensitive information before translation or external processing ECO orchestration Operate translation, document processing, retrieval, APIs, and governance in one environment Evaluation and AI QA Measure performance, analyze errors, test regressions, and validate workflow behavior Workflow design Define quality thresholds, review policies, terminology sources, privacy, and deployment requirements
Intelligent publishing
Automate publication by policy rather than by blind confidence in a model
The objective is to apply human expertise where evidence, risk, or quality requires it while allowing dependable content to move automatically.
01
Classify the content
Identify language, domain, document type, audience, sensitivity, and publication risk.
02
Select the workflow
Choose the appropriate engine, adaptation assets, terminology, privacy controls, and document-processing route.
03
Evaluate the result
Apply MTQE, terminology checks, formatting validation, and other quality signals before release.
04
Publish or escalate
Release content automatically, correct it, request human review, or stop publication according to policy.
Privacy-controlled architecture
Enterprise adaptation must not require surrendering enterprise knowledge
Translation memories, terminology, contracts, technical documentation, internal procedures, and reviewed corpora can contain commercially or legally sensitive information.
Privacy-controlled translation separates the value of those assets from uncontrolled external exposure. Retrieval can be limited to approved sources, processing can occur in private environments, and only authorized components receive the necessary context.
The same principles support private cloud, on-premises, air-gapped, and sovereign deployment strategies where organizations require greater control over infrastructure, models, data flows, and logging.
SOURCE
Controlled knowledge sources
Limit retrieval to approved terminology, memories, reference documents, and organizational data.
MASK
Anonymization before processing
Detect and protect personal or sensitive information before content enters translation or AI workflows.
DEPLOY
Private infrastructure
Operate services through controlled SaaS, private cloud, on-premises, or isolated deployment models.
AUDIT
Traceable decisions
Record model selection, retrieved evidence, quality scores, corrections, review actions, and publication outcomes.
Research outcomes
The project connected adaptation research with production-quality translation control
Its lasting value lies in the combined architecture rather than in a single isolated model or benchmark.
RAG
Controlled retrieval
Integration of selected bilingual and domain information into translation generation.
ADAPT
Dynamic adaptation
Stronger use of client terminology, approved language, reviewed material, and contextual evidence.
MTQE
Automatic quality estimation
Reference-free scoring to support production quality gates and targeted review.
APE
Automatic post-editing
Corrective processing designed to improve output before or instead of human intervention.
AGENT
Agentic workflow control
Structured decisions around retrieval, translation, verification, correction, and escalation.
ECO
Operational integration
A path from research components to document workflows, APIs, privacy services, and enterprise deployment.
Public evidence and current capability
Follow the path from CDTI research to Pangeanic’s production translation stack
These resources document the project, its current commercial continuation, and the quality and orchestration layers that make automated multilingual publishing operational.
Official project completion article
CDTI and ERDF Research Project
Review the published funding figures, project duration, research objectives, and relationship with Deep Adaptive AI Translation.
Read the project article → Adaptive translation productDeep Adaptive AI Translation
Apply translation memories, terminology, style, domain resources, and contextual retrieval within a controlled enterprise workflow.
Explore adaptive translation → Automated quality controlMachine Translation Quality Estimation
Use reference-free quality scores to automate release, corrective processing, targeted review, and rejection.
Explore MTQE → Operational platformECO Intelligence Platform
Operate translation, document processing, anonymization, MTQE, multilingual retrieval, and APIs in one governed environment.
Explore ECO → Complete R&D portfolioResearch and Innovation Projects
Place this CDTI project within Pangeanic’s Spanish, regional, and European research trajectory.
Explore the complete portfolio → Direct European programsEuropean AI and Language Technology Projects
Explore Pangeanic’s directly funded European work in translation infrastructure, multilingual data, privacy, cultural heritage, and language equality.
Explore European projects →Continue exploring
From multilingual research to governed production systems
Explore the wider project portfolio and the research programs that contributed data, models, routing, privacy, and multilingual infrastructure to Pangeanic’s current platform.
Complete research portfolio
Research and Innovation Projects
Explore Pangeanic’s European, Spanish, and regional projects across multilingual data, translation, privacy, model alignment, and sovereign AI.
View all projects → European research recordEuropean AI and Language Technology Projects
Review the directly funded European projects that created reusable language data, translation models, public infrastructure, and privacy technologies.
Explore European projects → Neural translation infrastructureNTEU
Discover the Pangeanic-led project that created direct neural translation models across the official EU languages.
Explore NTEU → Translation orchestrationiADAATPA and MT Hub
Explore the earlier European project that connected public administrations with multiple translation engines through secure routing and common APIs.
Explore MT Hub →Adaptive translation and intelligent publication
Build a translation workflow that knows what to publish, what to correct, and what to send for expert review
Pangeanic connects Deep Adaptive AI Translation, MTQE, terminology, privacy controls, document processing, and human review through the ECO Intelligence Platform.