Portfolio
×
Description
My MSc dissertation at Imperial College (2023-24) involves using
state-of-the-art generative AI frameworks to learn population-level patterns
in the shape of cardiac components. Analysis of these shapes are crucial for
reliable diagnosis of heart diseases. A snippet of my abstract is provided
below.
"Cardiovascular diseases (CVDs) cause over 20 million deaths annually, with
a third occuring prematurely in people under the age of 70. However, CVDs
are largely preventable with early detection and intervention. Over recent
years, there has been rapid progression in the development of automated
techniques for cardiac magnetic resonance imaging (MRI) analysis. Accurate
delineation of cardiac components is crucial to assist in anomaly
detection and diagnosis, and shape analysis is an essential
prerequisite.
I am very grateful to Prof. Elsa Angelini and Prof. Loïc Le Folgoc for
supervising me.
The emergence of deep learning has introduced powerful frameworks capable
of automating the process of learning compact shape representations.
Variational autoencoders (VAEs) are a class of generative models that
excel at learning efficient low-dimensional representations of complex
data. In particular, the Nouveau VAE (NVAE) is a deep hierarchical VAE
that is the state-of-the-art among its class in encoding fine-grained
details in high-resolution images.
In this dissertation, we examine how the NVAE framework can be applied to
cardiac shape analysis..."
Source code will be available soon.
×
Description
Computer-aided diagnosis (CADx) is very important in clinical environments
as they assist radiologists in interpreting medical images. This involves
early detection of lesions, as well as enabling more accurate diagnosis.
AI in general is advancing very rapidly. So, can we use these new,
state-of-the-art AI techniques and integrate them within CADx?
My BEng dissertation at Imperial College (2022-23) focuses on evaluating a
specific framework for medical imaging: SimCLR. SimCLR is a self-supervised
contrastive learning framework, which enables pretraining large
convolutional networks like ResNet with unlabelled data. As a result, only a
small amount of labelled data is required for downstream tasks like
finetuning the model for classification. Traditionally, supervised learning
requires lots of labelled data, so this is a very exciting prospect,
especially since clinical environments tend to lack labelled data. The
thesis involves proposing and evaluating novel techniques and workflows for
medical images that improveme over existing SimCLR methods.
I am very thankful to Prof. Ben Glocker for his supervision and guidance.
Source Code
×
Description
In Spring 2024, we (group of 2) developed an ensemble model (transfer
learning of pretrained RoBERTa and DistilRoBERTa models) to detect
Patronizing and Condescending Language as described by
SemEval 2022 Task 4: Subtask 1.
Our model achieved 0.5548 F1 score on the test set, which would've ranked
22nd on the
contest. (The top performing teams were industrial research groups.)
It outperforms baseline RoBERTa by over 0.06 in F1 score.
This project was the coursework for the Natural Language Processing module
at Imperial College, for which we achieved full marks.
Source Code
×
Description
This lightweight Python application allows you to render or record highly customisable bar chart animations from CSV files, like those data visualisation videos on YouTube.
Source Code
×
Description
Duel! is a 2P shooter game I developed in Unity in 1 week for Brackeys Jam
2020.2.
Given the time constraints, I opted for a simple game idea that is easily
extendable, with a focus on providing an elegant interface and user
experience. This involved smooth animations, a custom keybinds option, as
well as rewriting the player physics, as Unity's default physics engine was
rough and inconsistent. Overall, it was very exciting to work with limited
time.
Watch the "behind the scenes" video:
https://www.youtube.com/watch?v=2NnTHHqoIrw
Try out the game! Grab a friend, and duel to the death!
Preview
×
Description
Ten Minutes of Existence is a puzzle / adventure game I developed for the
GameDev.tv Community Jam - May 2020.
This was a fun project during the initial lockdown, especially working with
Shader Graph and the Universal Render Pipeline to texture and animate
rippling water.
Watch the "behind the scenes" video:
https://www.youtube.com/watch?v=xrW_MobboxE
Walkthrough (spoilers!):
https://www.youtube.com/watch?v=f8kdB0yYRlE
Preview
×
Description
Peer Assessment is a web app used by the Department of Computing at Imperial
College, allowing students to peer-assess each other and providing staff
with a seamless administration interface. I was one of the main developers
during my internship with DoC EdTech, Imperial College in Summer 2021.
I am very thankful to Ivan Procaccini, Dr. Andrea Callia D'Iddio and Dr.
Robert Chatley who supervised me and introduced me to test-driven
development and other modern agile development practices.
After my internship, I volunteered to develop the official site for Doc
EdTech at
https://edtech.pages.doc.ic.ac.uk
Source Code
×
Description
With the prevalence of autoregressive generative models being made
publically available, in Autumn 2022 we (group of 6) created a distributed
framework for automated gamebook generation using OpenAI's GPT-3 model.
Create Your Own Adventure (CYOA) offers an extensive set of tools to allow
authors to easily create, fully customise and print their gamebooks.
Frontend:
https://github.com/j-freddy/icl-cyoa-frontend
Backend:
https://github.com/j-freddy/icl-cyoa-backend
×
Description
WACC is an imperative language that supports fundamental constructs such as
variables, loops, conditional branching, arrays and functions. The preview
image is a code snippet from a Tic-Tac-Toe AI written in WACC.
Our challenge in the 2022 Spring term? Given 6 weeks, write a compiler that
performs syntax and semantic analysis on any WACC file, then (if valid)
generates executable code for the ARM11 architecture. The four of us
successfully implemented this in Scala and Parsley (parser combinator
library) and got a 99% score in the backend code review.
×
Description
SolarSearch is powered by a deep learning neural network AI and developed
during IC Hack 22 in 24 hours.
Given a country, the web app generates a ranking of areas within the country
that are most feasible for building solar farms. The ranking algorithm
accounts for terrain, weather, atmospheric pollution and other factors,
using a mixture of computer vision and third-party data.
More details can be found here:
https://devpost.com/software/solarsearch
Source Code
×
Description
In Summer 2021, we (group of 4) created a two-pass assembler that converts ARM assembly code into binary. We also coded an emulator that runs binary code, complete with the three-stage pipeline, CPSR flags and barrel shifters.
×
Description
I have always been an avid enjoyer of platformer games, dating back to when
I discovered SuperTux in primary school. This is my attempt at developing
one in Unity, and I was quite happy with how it turned out.
Unfortunately, I never turned this into a fully-fledged game. You can play
the demo using the link below.
Preview
×
Description
Pintos is an operating system with a monolithic kernel. It supports basic
functionality such as multithreading, running user programs and file
systems.
Our challenge in Fall 2022 is to implement and extend its functionality,
including non-busy wait, implementing system calls, priority scheduling and
priority donation.
×
Description
During my 1st year at Imperial, the Department of Computing held an AI
Kotlin Pawn Race, and I was one of the winners! (using a MCTS strategy)
Fast forward to the Summer term and I was
programming chess
in Java as a side hobby. I prototyped it to support Pawn Race and challenged
myself to create a stronger AI. This time, I used Minimax with alpha-beta
pruning, using the strategy of "space control" as a static evaluator. It
triumphed over my Kotlin AI.
Source Code
×
Description
This is my project for the AQA A-level Computer Science NEA. Back then, I already had a deep interest in machine learning, and used the opportunity to research and build (from scratch) a genetic/evolutionary algorithm to train AIs to play a platformer.