"Desarrollo de una tecnología de IA multimodal basada en LLM para la gestión integral de datos" A research and development project to automate and improve enterprise data management processes.
The challenge
Our team aimed to create a multimodal AI technology based on NLP and LLMs, capable of:
- Extracting information from unstructured and semi-structured sources
- Validating and improving data quality automatically
- Transforming data through natural language processing
- Providing advanced analytics and explainable insights through LLMs and RAG architectures
For this project, we received funding through the PID programme (Proyectos de Investigación y Desarrollo, a funding initiative developed by CDTI (Centro para el Desarrollo Tecnológico Industri) to support business R&D projects.
This funding enabled us to begin working on the solution.
What we did
Our team built the solution on four layers:
- A multimodal extraction layer from unstructured or semi-structured sources using advanced NLP and LLM technologies.
- An automated data validation layer designed to support more robust, higher-quality data.
- A governed NLP transformation layer, parameterizing the LLM to understand and execute transformation commands in a way that is accessible to non-technical users.
- A RAG + LLM infrastructure that combines retrieval and generation to provide tabular data analysis without the need for technical intervention.
Together, these layers support the complete data management lifecycle in six stages:
- Data ingestion: PDF documents, Excel files, free-form text...
- Data extraction: OCR, layout analysis, LLM-based extraction and schema enforcement.
- Data validation: Deterministic, semantic and confidence-based validation with human review.
- Data transformation: Natural-language-to-SQL conversion on governed data models.
- Data analysis: RAG-powered document retrieval and analytics on structured data.
- Explainability: Evidence-based responses with citations, traceability and controlled non-response.
To validate the approach, four proof-of-concepts were developed and tested using real-world scenarios: clinical guideline analysis, vascular surgery report processing, meeting transcript intelligence and a natural-language SQL chatbot.
What we achieved
The project delivered a modular and reusable AI architecture covering the complete data lifecycle, from extraction to analytics, with clear, traceable communication for users.
- Hybrid wins: using an LLM + deterministic approach is the most robust approach
- Validating is not a single mechanism: each type of data requires its form of control.
- What sounds right is not always evidence-based: an answer can appear valid and fluent without being supported by the provided evidence, so validation for this kind of solutions must be systematic
- Governance is indispensable: governance and control actions must be at every stage of the cycle, not only during generation
- Evidence retrieval and synthesis must be separated: evidence acquisition must remain separate from evidence synthesis, with full traceability
- A transferable approach: the combination of supervised adaptation, dynamic context and parameterized prompts reduces costs and brings this type of technology into specialized domains.
The project "Developing a multimodal LLM-based AI technology for the full data management lifecycle" ("Desarrollo de una tecnología de IA multimodal basada en LLM para la gestión integral de datos") was funded by CDTI.

.png)