Hello, I'm Eduard Ferré

Built a multi-agent AI system (Python, OpenAI, Azure) that traces data lineage across ~80,000 COBOL/JCL Mainframe files — around 24.5M lines of code — for a large European financial-automotive group. Sole owner for the first 8 months, from architecture to delivery and executive-level client contact; now co-leading a team, with regulatory deliverables compliant with ECB requirements.

Architected a Knowledge Graph + RAG codebase-intelligence platform for a banking and insurance group, adopted by most of its technical and business teams: a multi-repository graph (COBOL/JCL, Java, JS/TS, Python) built via AST parsing (ANTLR4), enriched with LLM descriptions and business context, indexed in Neo4j and queried with GraphRAG (BM25 + embeddings, neighbour expansion, LLM traversal through tool calling). Also built a guardrailed RAG chatbot for an international insurer's website that routes users to the right section (in UAT, production launch planned).

Designed a LiteLLM-based AI governance proxy shaped as a product for external sale.

Other work: a sales-training pipeline, an accounting application for a Spanish public administration body, and client PoCs.

Developed key modules for a multi-tenant remote eSIM/eUICC profile management platform (SGP.22, SGP.32) used by multiple telecom operators — implemented terms-of-use/privacy acceptance flows and an orchestration proxy for dynamic eSIM profile recycling and pooling across devices, enabling massive-IoT use cases such as Spain's DGT V16 emergency beacon. Extracted and audited user data to support the legal department.

Backend in Java (Spring Boot) and TypeScript (Node.js/Express), frontend in Angular, over Oracle SQL and MongoDB — working Scrum/Agile with Jenkins for CI/CD and SonarQube as the quality gate.

Research and characterization of different IPV (indoor photovoltaic) cell technologies under varying indoor light conditions, as part of the EU-funded project CELISA — efficient harvesting of indoor light energy for autonomous sensors.

Owned the full experimental cycle: circuit design and instrumentation (KiCad, BenchVue, AvaSoft), measurement campaigns under controlled illuminance and temperature, and data analysis in Matlab — work that ended up in four peer-reviewed publications.