My research is at the intersection of Deep Reinforcement Learning
and
Large Language Models, with a focus on the
Credit Assignment Problem.
I am particularly interested in how AI evaluates actions, and in scaling reinforcement learning
to
enable
open-ended learning and real-world
generalisation.
University College London
Assistant Professor (UK Lecturer, Teaching)
September 2024 – Present
London, UK
I teach the Data Acquisition and Processing Systems within the Integrated Machine Learning Master at UCL. The module is an introduction on how to structure a machine learning project: data acquisition using web API and sensors; data storage, SQL and noSQL databases; statistical data processing including sampling, normalisation, and linear projections; deep learning. Secondly, I am responsible for the coordination of the student’s final thesis project across the department. Finally, I supervise postgraduate students in their final projects.
Imperial College London
Research Assistant
September 2019 – November 2020
London, UK
I conducted research to shorten the computational time for predictive modelling of surgical interventions in cardiology. Statistical modelling, specifically deep learning, is central in the approach. The deep networks take advantage of accepted numerical modelling techniques to generate training data and are optimised to infer approximate solutions in about 1% of time necessary to standard models. The position was founded by the Rosetrees Trust, in collaboration with the ElectroCardioMaths program, a multidisciplinary initiative that brings together the National Heart and Lung Institute, and the Departments of Bioengineering, Aeronautics, Computing and Physics to address key challenges in the diagnosis and treatment of complex cardiac arrhythmia.
BuroHappold Engineering
Machine Learning and Decision Analytics Lead
August 2018 – October 2020
London, UK
I led the applied research in Machine Learning to help understand how AI can create value for the business. We democratised the access to Deep Learning technologies to allow every employee to access the knowledge and the tools. We exploited Visual Programming, a recognised and diffused tool for design, to create a framework that interoperates between the most common deep learning libraries, tensorflow, keras, pytorch, numpy. The position was part of the wider computational core team that brings together discipline leads into a centralised research team bhom.xyz. The BHoM is currently adopted in different companies and is at the base of the project funded by Innovate UK, github.com/aecdeltas, which I advise for.
BuroHappold Engineering
Computational/Machine Learning Engineer
August 2017 – August 2018
London, UK
For 50% of my time, I have been applying deep learning for computer vision to the analysis of security footage for the Premier League. To monitor the number of standing fans during a football match we created a database of more than 400,000 annotated images and trained a convolutional deep network to identify them. We then exploited principal component analysis, hierarchical clustering and bespoke data visualisation to gather insights from the resulting probability distribution. For the remaining 50% of my time, I have been designing a framework for data sharing and co-creation in design, for the architecture, engineering and construction industry, bhom.xyz.
We created a software-agnostic model to link together the capabilities of existing software and allow seamless interoperability between them. A short-cycles, distributed scrum development model, and an entity-component-system architecture allowed independent contributions from more than 50 users. My main responsibility has been to lead the UI and support the framework leadership of the project.
BuroHappold Engineering
Intern Computational Engineer
April 2017 – August 2017
Bath, UK
I provided computational support for the Stadia Atmosphere project and helped introduce deep learning into the current offer for sports venues design. We used pre-trained deep neural networks for Natural Language Processing to perform sentiment analysis on news regarding a specific football team.
Gianni Ranaulo Design
Architect
September 2016 – December 2016
Dubai, UAE
I provided support for the parametric modelling of a façade in a multi-purpose shopping centre.
Gridshell.it
Computational Architect
September 2015 – September 2016
Naples, Italy
I conducted research on the application of computational tools to recover the use of traditional low-tech construction techniques. I used generative modelling to provide cost-effective, environmentally efficient, and functionally viable structure. Using genetic algorithms, particle-spring system models and dynamic relaxation we designed and built 13 prototypes of timber post-formed gridshells. Taking advantage of recognised acoustic modelling techniques, we generatively designed three temporary acoustic shells for outdoor classical concerts, the last of which has won the Peter Lord Award.
Gridshell.it
Computational Architectural Assistant
July 2014 – September 2015
Naples, Italy
CRC – Constructions Restorations and Consolidations
Intern Architect
September 2012 – December 2012
Naples, Italy
I provided support for the preparation of compliance documentation for a multi-storey parking building. My main responsibility was to ensure the fire compliance of the building.