López de Hoyos Street 64, 1G. 28002 Madrid, Spain
Advanced Conversational Data Analysis Platform
Sumiriko
Generative AI, Multi-Agent Architecture, Data Analysis & RAG
We designed a solution that transforms how organizations query and exploit their critical business information. Through a generative AI model, we developed a conversational platform that allows any user to obtain complex insights through natural language, eliminating technical barriers and significantly reducing response times.
Our specialized multi-agent approach not only answers questions about SQL databases, but integrates capabilities for generating graphs, document analysis and file processing (CSV, Excel, PDF). All of this deployed in an On-Premise environment that ensures data confidentiality, with hybrid architecture prepared to scale to the cloud as needed, always maintaining total control over sensitive information.
Challenges
- Specialized multi-agent architecture: Design of a system router (Planner) that directs queries to specialized assistants (SQL, Graphics, Documental Analysis), optimizing the accuracy of responses according to the context of each question.
- Automatic generation of SQL and visualizations: Implementation of agents capable of translating natural language into valid SQL queries against SQL Server, automatically generating interactive graphs when the question requires visual representation.
- Fluid integration with Open WebUI for conversational experience: Development of customized pipelines that connect LangGraph with Open WebUI, implementing asynchronous processing using Redis queues to maintain the responsive interface during long-duration operations. Management of streaming responses, conversational state and rendering of graphs in chatbot format accessible to non-technical users.
- On-Premise security and control with hybrid scalability: Deployment of local LLM models ensuring confidentiality, with architecture prepared for Azure resource consumption without compromising data privacy.
Key Success Factors
Multi-agent orchestration with LangGraph
Strategic selection of specialized LLMs
Intelligent context system based on YAML
Automatic generation of visualizations with volume validation
Methodology
Iterative proof of concept
We started with a PoC focused on the basic SQL Assistant, validating the ability to generate queries before expanding to multiple specialized agents, ensuring technical feasibility in each phase.
Design Thinking applied to agent architecture
We identified usage patterns through workshops with end users (business analysts), designing specialized assistants according to real use cases: direct queries, visual analysis and documentation.
Modular development with reusable subgraphs
We implemented reusable components (sql_subgraph, graph_subgraph) that encapsulate common logic, allowing the composition of complex flows without code duplication and facilitating independent testing.
Synthetic data testing and domain validation
We create test datasets based on real YAML schemas, validating SQL and graph generation before connecting to production databases, reducing risks of sensitive data exposure.
Gradual deployment with monitoring
We implement first in controlled On-Premise environment with extensive logging of LLM-user interactions, allowing prompt refinement and early detection of generation failures before opening to end users.
Results
- The developed solution allowed centralizing all critical business information and accessing it through a conversational assistant, significantly reducing the time required to obtain relevant insights. Thanks to the AI model, users without technical knowledge can query complex databases and receive complete analyses in seconds, replacing manual processes that previously required hours. Automation of document processing and immediate generation of graphs and summaries improved operational efficiency and accelerated decision-making.
- The platform, deployed in a secure On-Premise environment and with hybrid architecture, guarantees total control over sensitive data while allowing the computation to scale when necessary.