Leonardo Figueiredo

LuminLab: An AI-Powered Building Retrofit and Energy Modelling Platform

📍 PLEA 2024 Conference 🏛️ Trinity College Dublin & Maynooth University 📅 2024 🤖 AI/ML, Energy Modeling

This research presents the development of LuminLab, an innovative AI-powered platform that revolutionizes the building retrofit process. The platform integrates a purpose-built human-centric AI chatbot with predictive energy modeling to provide streamlined, natural language-driven retrofit planning for homeowners.

Key Contributions & Technical Implementation:

AI Integration: Fine-tuned Large Language Models (LLMs) serve as intelligent interfaces for retrofit guidance
Full-Stack Development: React and Material-UI frontend with integrated energy modeling backend
Pragmatic Design: De-silos stakeholder knowledge and improves communication between homeowners, contractors, and energy assessors
Cost Optimization: Addresses the €600-€800 energy audit barrier by providing on-demand retrofit planning
3D Visualization: Planned integration with Neuralangelo technology for architectural modeling from home videos

Impact: The platform addresses critical challenges in Ireland's building retrofit landscape, where operational building energy use represents 30% of global consumption. By streamlining the traditionally complex and costly retrofit process, LuminLab empowers individual homeowners to make informed decisions about incremental improvements that contribute to Ireland's net-zero emissions targets.

My Role: As a co-author, I contributed to the technical development of the AI chatbot system, the integration of machine learning models for energy optimization, and the overall platform architecture design.

📄 Download PDF

Multi-method Phenotyping of Long COVID Patients Using High-dimensional Symptom Data

📍 Under Review - Nature 🏛️ Patient-Led Research Collaborative 📅 October 2024 🧬 Machine Learning, Bioinformatics

This groundbreaking research addresses one of the most pressing medical challenges of our time: understanding Long COVID's complex symptom presentation. Working with the Patient-Led Research Collaborative on a project funded by Ethereum founder Vitalik Buterin, our team analyzed the most comprehensive Long COVID patient dataset to date.

Technical Methodology & Findings:

Multi-Method Approach: Applied three distinct clustering algorithms - Autoencoders with HDBSCAN, Ensemble Clustering, and Latent Class Analysis
High-Dimensional Analysis: Processed 162 symptoms across 6,031 patients with comprehensive temporal tracking
Genetic Algorithm Optimization: Used evolutionary computation to optimize neural network hyperparameters for symptom clustering
Robustness Assessment: Evaluated clustering stability using Adjusted Mutual Information (AMI) across different methodologies
Clinical Insights: Identified recurring patterns including age and gender correlations with symptom clusters

Key Discovery: While all three methods produced clinically plausible symptom clusters, concordance across methods was surprisingly low. This finding has profound implications for Long COVID research, suggesting that the complexity of symptom presentation may be easily missed by overly simplistic clustering approaches.

Research Impact: Our work highlights the importance of multi-method validation in medical machine learning and provides a roadmap for future Long COVID phenotyping studies. The research emphasizes the need for semi-supervised approaches and additional patient data beyond symptom reports.

My Role: As one of three co-first authors, I led the development of the autoencoder-based clustering methodology, implemented the genetic algorithm optimization pipeline, and contributed to the comparative analysis across all three clustering approaches.

📄 Download PDF