Design and deployment of production‑ready LLM systems for hospitals and public institutions (RAG, agents, real‑time voice).
RAG Systems
Implementation of retrieval-augmented generation systems combining knowledge bases with LLMs for precise and contextual responses.
ReAct Agents
Development of intelligent agents that reason and act autonomously, integrating thought and execution in iterative loops.
Fine-tuning
Fine-tuning of base models (GPT, LLaMA, Mistral) for specific tasks, optimising performance in specialised domains.
Prompt Engineering
Advanced prompt design with techniques like chain-of-thought, few-shot learning, and self-consistency to maximise LLM effectiveness.
Featured Projects
🏥 ASISVIA — Case study (Hospitales Universitarios San Roque, CDTI PAIS-20241015)
- Context / Client: Hospitales Universitarios San Roque — patient experience surveys in production (CDTI project)
- Problem:
- Need scalable, empathetic patient interviews with automated mood & satisfaction analysis
- Reduce manual post-hoc processing of free-text/voice survey responses
- Solution / Architecture:
- LangGraph-based StateGraph conversational workflow with SessionContext for isolated multi-room management
- LiveKit WebRTC audio streaming → Whisper STT → GPT-4o-mini for empathetic logic → ElevenLabs TTS for natural voice
- Hybrid templated & generative modes to control cost and behaviour (templated = deterministic, low-cost; generative = adaptive follow-ups)
- Voice analysis module for real-time satisfaction/mood extraction + PostgreSQL checkpoints for audit & resume
- Impact / Status:
- Supports concurrent multi-room sessions in hospital environments; deployed for patient experience surveys
- Designed to reduce manual analysis and provide timely emotional‑wellbeing insights to clinical staff
🏙️ CityVerse — Case study
- Context / Client: Supports EU officers evaluating urban designs inside VR environments
- Problem:
- Need structured, repeatable evaluation workflows for VR-based design reviews
- Ensure objective, comparable feedback across reviewers and sessions
- Solution / Architecture:
- Multi-stage structured evaluation workflow orchestrated by a finite state automaton
- CityBot agent (GPT-4o) for domain-guided questioning and summarisation
- Socket.IO real-time integration with Unity/web clients + ElevenLabs TTS for voice interaction
- Result export for comparative analysis and policy reporting
- Impact / Status:
- Enables consistent, auditable VR-based evaluations used by decision-makers
- Prototype validated in pilot sessions; designed for scale and repeatability
Work Methodology
- Requirements Analysis: Precise identification of problem and project objectives
- Architecture Design: Selection of appropriate models, vectorstores, and frameworks
- Iterative Implementation: Development with continuous testing and prompt adjustment
- Evaluation: Performance metrics (accuracy, latency, costs) and reliability (error handling, logging, monitoring)
- Deployment: Production deployment with observability and incident response procedures
Featured Research
📚 OntoGenix: Leveraging Large Language Models for Enhanced Ontology Engineering from Datasets
Semi-automated system utilising LLMs for ontology engineering from datasets. Published in Information Processing & Management, 2025.
View full publication →