Python (Pyomo, Pandas, NumPy), React.js, Node.js, Express, Socket.io, MySQL, Docker, Terraform, AWS, GCP

Led the development of a software suite which provided consultancy support and strategy advice to multiple leading global renewable energy power generators (such as Corio and Invenergy). Developed tools which simulated auctions and forecasted results based on in depth competitor analysis and market evaluations, as well as tools which helped clients navigate the complex auctioning mechanisms and learn principles of game theory enabling successful results.

Discusses how AI can be used and implemented to address the various challenges faced in the wind energy sector and the wider ethical implications of using AI in this domain.

Python (OpenCV, Matplotlib, NumPy, Pandas, TensorFlow, scikit-learn, Seaborn)

Single-image super-resolution (SR) is an important computer vision challenge that aims to reconstruct high-resolution images from low-resolution ones. Various statistical algorithms were employed for this purpose until deep learning models could be used to address this in a much more efficient manner. This report1 details the research and deep learning models that were used to evaluate their SR performance against benchmark datasets, using common image quality assessment (IQA) metrics such as peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) score.

The goal of single-image super-resolution (SR) is to reconstruct a high-resolution (HR) image from a low-resolution (LR) image. It is a classical inverse problem in computer vision that is undetermined in nature, with no single unique solution. This process is incredibly useful in situations where HR images are required for analysis, but only LR or low-quality images are available; such as medical diagnosis, satellite images, surveillance, astronomical images, and video analysis and enhancement. Super-resolution has widespread application in a multitude of industries. The increase in research and development in this field in the last few years has made this an extremely challenging and complex task efficient and improved its quality.

There are various techniques that can be used for image super-resolution. These range from classical interpolation-based statistical algorithms to deep learning approaches which employ models such as convolutional neural networks (CNNs) and generative adversarial networks (GANs). This assignment aims to assess some of these single-image SR models using popular benchmark datasets for training and evaluation.

This report summarises the processes followed to evaluate the performance of the single-image super-resolution deep learning models like SRCNN, FSRCNN, and ESRGAN models against benchmark datasets and classical algorithms, on various scaling factors. Although one of the models trained in this project achieves results comparable to industry standards, computing hardware limitations restricted the performance of the trained models. Thus, in order to improve the quality of models trained in this assignment, a much larger number of generated patches (with rotations) need to be used. Despite the adaptability and versatility of state-of-the-art models used in this assignment, there are still scenarios in which the perceptual quality of upscaled images is lacking. Furthermore, most industry standard models are trained on images downscaled via bicubic interpolation. The applicability of extending this to other method needs to be studied for its potential to increase the models’ adaptability to distortions and noise.

Python (Matplotlib, Pandas, scikit-learn, Seaborn, Prophet, Scipy, yFinance), MongoDB

This assignment covers the process of predicting American Airlines stock prices, as well as attempting to improve the quality of prediction using various auxiliary data sources, such as data for the COVID-19 pandemic, passenger volumes, crude oil fuel prices and weather data. The experimental results show that it is possible to improve the performance of forecasting models using external data sources, but its effects are unpredictable.

Stock market predictions are important as they can help investors make informed decisions about buying or selling shares in a company. They can also be used by traders to make short-term predictions about the direction of the stock market. Additionally, stock market predictions can be used by analysts to assess the overall health of a company and its industry.

This project aims to predict the stock prices for American Airlines (AAL) for May 2022, using data collected from the start of April 2017 to the end of April 2022. In addition to predicting the stock value exclusively using the data provided with the stock, the project also attempts to acquire, process, explore, and analyse other data sources, that could be used to improve the quality of this prediction. Due to the time period covered in this project, the COVID-19 pandemic becomes an obvious choice for an external data source, especially considering the impact it had on the airline and travel industries. Additionally, the impact of passenger volumes, weather, crude oil prices and occurrences of US Federal holidays were also considered.

This report summarises the processes followed to predict American Airlines stock's adjusted closing prices for May 2022, using stock data (along with auxiliary data) from April 2017 to April 2022. Data such as stock indexes, technical indicators, COVID-19 pandemic cases, crude oil fuel prices, US airport passenger volumes, and weather data are used to explore relationships with the stock value and underlying patterns, in an attempt to improve the performance of the forecasting model. Although this project achieves good results, further progress can be made by acquiring and experimenting with additional sources of auxiliary data, such as GDP and inflation rates in the US, along with more specific statistics (like load factor, ticket prices, number of flights, etc.) for American Airlines. Additionally, transformations can be applied to the dataset in order to use other models.

Python (Parselmouth (Praat), Pandas, scikit-learn, TensorFlow, SpeechRecognition, PyCryptodome), AWS

Smart home speakers have gained widespread popularity over the last few years and have become a common item in households. Due to their access to sensitive speech data, it is important to evaluate the threats prevalent in the smart speaker environment. This project aims to analyze threats such as third-party data access and data exposure and implements defense methodologies to evaluate their effectiveness against these threats without hindering the required functionality.

A smart home speaker is a device that allows you to interact with various applications and other smart home appliances in your home, via voice commands. Through voice recognition, these devices allow you to give complex commands to other smart home appliances (such as thermostats, lights, doors, etc.) as well as retrieve information from various sources (such as weather, traffic, news, etc.). Many smart home speakers also have built-in virtual assistants, such as Amazon’s Alexa, Google Assistant, or Apple’s Siri, which can help you with various tasks such as setting reminders, answering questions, and making phone calls. Smart home IoT (Internet of Things) devices such as smart speakers and screens have become extremely common in households. Market penetration for smart home speakers has more than doubled since 2017, increasing from 7.2% to a forecasted 16.7% by 2023. 70% of users who use voice assistants to interact with smart devices use smart speakers to interact with them.

In order to provide this functionality, these devices are connected to the internet and often send voice recordings of the commands given to the service provider’s servers/cloud networks via the internet (Fig 1). This is done in order to perform real-time speech-to-text (STT) to process the command, but also additional functions such as contextual understanding and improving the performance of voice recognition capabilities.

In this project, we aim to identify and assess various security and privacy threats to the smart speaker environment and implement and evaluate the effectiveness of defense mechanisms to mitigate these threats.

This project identifies security and privacy threats to the smart home speaker environment and evaluates the effectiveness of security measures taken to address these threats. The paper proposes a combination of audio transformation and end-to-end encryption to mitigate the threat of data exposure and third-party data access. This project demonstrates that it is possible to decrease the effectiveness of ML-based models used for data extraction by masking characteristics in the audio files. This process can greatly increase privacy, while only marginally affecting the performance of the speech-to-text models. We have also demonstrated that end-to-end encryption can be used on audio files with no loss of speech information while greatly increasing the security and privacy of audio files.

Python (OpenCV, NumPy, Pandas, scikit-learn, TensorFlow, Pillow, Matplotlib, imutils)

Image classification and facial recognition is a challenging, yet widely used problems in machine learning. It has become a part of our day-to-day lives; from face unlock technologies to camera autofocus and QR code scanning. Thus, it is important to explore the various ways of implementing this technology and to find ways to improve its efficiency.

This project aims to explore various combinations of state-of-the-art feature extraction and classification methodologies to address the image classification problems posed in this assignment. The CelebA and Cartoonset datasets are used for these tasks, which involve gender detection, emotion detection, face shape recognition and eye colour recognition. In order to perform these classification tasks feature extraction techniques like facial landmark detection and embedding are used along with classifiers like support vector machines, convolutional neural networks, k nearest neighbour classifiers and random forest classifiers.

This report aims to explore various combinations of state-of-the-art feature extraction and classification methodologies to address the image classification problems posed in this assignment. The CelebA and Cartoonset datasets are used for these tasks, which involve gender detection, emotion detection, face shape recognition and eye colour recognition. In order to perform these classification tasks feature extraction techniques like facial landmark detection and embedding are used along with classifiers like support vector machines, convolutional neural networks, k nearest neighbour classifiers and random forest classifiers.

During this assignment, various combinations of feature extraction processes and classification methods were used. Feature extraction is an extremely important step in the process for machine learning tasks in general, but especially for image classification tasks. Without proper feature extraction techniques, too many variables and noise are introduced to the model, which can help with the regularisation of the model, but more often than not, can result in the learning of unwanted features.

In this assignment, dlib, MTCNN and HaarCC facial detectors were used for feature extraction. Even though state-of-the-art tools are able to accurately detect faces (and other features) in a large proportion of unseen images, they are not able to do so for all of them. It was expected that this problem would be observed with the cartoonset dataset samples, but it was also observed with CelebA samples. dlib had poor performance against the cartoonset dataset; not only was it not able to detect features in about 20\% of the samples, but even the detected features were not accurate. On the other hand, HaarCC had slightly better performance on the cartoonset dataset and was able to accurately extract relevant features. MTCNN was not tested against the cartoonset dataset. Thus, further improvements in feature extraction techniques are required to be able to perform such processes on any unseen image.

The models used in the tasks produced reasonable results, especially the Support Vector Machines and the Random Forest Classifiers. However, the results produced by the Convolutional Neural Networks were slightly underwhelming, given the potential that these deep learning algorithms have with image recognition. Thus, further work and research is required to improve the architecture and the effectiveness of the CNN used in this assignment.

Python, Azure

Android has become one of the leading mobile operating system platforms in the world. As of September 2020, there are over 3 million applications available on the Google Play Store. Considering that applications on Android are available via other sources as well (for instance, torrents, one-click hosters, websites, blogs or alternative app markets), the total number of applications available for the Android platform is much higher. In the android landscape, malware represents a significant proportion of all applications. Various sources estimate between 6-10% of android applications are malicious and potentially harmful

Due to constant development of new strains and variants of malware, the composition of malware found in realistic settings is invariably evolving. These changes can lead to inconsistencies in the performance of malware classifiers. It is infeasible to repeatedly retrain classifiers on newer strains of malware. This is due to the high cost of identifying and labelling the malware as well as the computational cost of retraining and distributing the updated classifier. Thus, there is a need to model malware classifiers which can stay effective against the evolving malware landscape. This can be achieved by identifying and studying program analysis methodologies in order to create robust abstractions of the applications.

Android is one of the world's leading tablet and smartphone operating systems. As of 2019, more than 1.6 billion devices worldwide support Android. which accounts for over 70\% of the mobile operating system share worldwide. Research shows that the number of total malware by 2020 is estimated at over 1.1 billion, which has increased 18 times over the last 10 years. Moreover, there have been an average of more than 10 million new malware per month in 2020.

As new forms of malware are being generated, it is important to assess whether models which have been trained on older data sets, can effectively and accurately classify newer applications as malicious or benign. An issue in the Android malware classification realm, is that malware classifiers are not evaluated in settings which represent real world deployments. In order to successfully assess the effectiveness of a classifier in a realistic setting, it is important to evaluate the robustness to temporal decay.

In addition to temporal considerations during classification, the ratio of malware to goodware (or non-malicious) in realistic scenarios should be enforced during evaluation. This ratio is domain specific. Reports suggest that most Android applications are non-malicious.

Thus, there is an urgent need for robust, efficient and spatial-temporal aware malware classification solutions.

The project is research oriented and focuses on evaluating various program analysis methodologies in the malware classification realm. To this end, static program analysis algorithms are considered for this project. The developmental scope of the project will cover the execution of control flow analysis of Android applications, generations of graph vector embeddings and modelling of a malware classifier using these abstractions. Furthermore, the project will evaluate the effectiveness of different abstraction methodologies for classification. With a focus on robustness of classification to changes in malware landscape over time, the project will also evaluate the classifier with time-aware data splits.

There are numerous static program analysis methodologies. However, this project will exclusively focus on control flow analysis and extensions thereof. The methodologies implemented in this project have been selected after careful consideration and are in no way an exhaustive list of approaches.

Due to the research oriented nature of the project, the development of a sophisticated malware detection system is not within the scope of the project. Even though the modelling and development of a classifier is within the scope of the project, the user interface and experience design for the same are not.

MongoDB, React.js, Node.js, Express, TypeScript, AWS

Responsible for developing the website and framework for a Singapore based logistics platform.

The website is under construction and will soon be published for use. Stay tuned for updates.

This coursework is set in a modern emulation of the 1980 maze chase arcade game of Pac-Man. In which the main agent (Pac-Man), explores the maze with the objective of consuming all the available food, while avoiding (or attacking) hostile (or frightened) ghosts. Near the corners of the maze are capsules which (on consuming) provide Pac-Man with the temporary ability to eat the ghosts, that is, change the states of the ghosts from hostile to frightened. Once Pac-Man consumes a ghost, the ghost returns to the "ghost home", where it is regenerated in its hostile state.

The aim of the project is to implement a MDP-solver to control Pac-Man's movements and win games in two different mazes (layouts). A win is defined as the state in which PacMan has eaten all the food on the map. A loss is defined by the state in which a hostile ghost has touched Pac-Man. For the project, the score of the game of is irrelevant; the aim is to win as many games as possible.

After various testing scenarios, I was able to determine the optimal values for the variables in the code. Using these values, I was able to get a win rate of 0.603 and 0.600 for smallGrid and mediumClassic respectively; thus, placing within the Epic and Legendary Pac-Man Skill Levels (according to the marksheet). However, due to the low number of runs during the marking process and randomness in win rates, the results during marking may differ.

Even though an Markov Decision Process (MDP) agent is able to produce high win rates for the Pac-Man environment, it is not consistent. There is a high degree of randomness and variation in this particular implementation of the algorithm. For consistent win rates (or even high scores), other algorithms and approaches (like ad-hoc ghost avoidance) should be explored.

React Native (JavaScript), Android, iOS, Microsoft Outlook

The My Buddy app is a fully functional cross-platform (Android \& iOS) mobile application specially designed for the Maudsley NHS Foundation Trust Helping Families Team which are a specialist mental health service for families in complex circumstances. It's a buddy system for the clinicians that allows them to keep track of their appointments, ensure their attendance and ensure the safety of the clinicians when they see their patients by having them check in/out at scheduled times, and notify members of the team if they haven't already. The entire app is centered around the objective to provide a seamless experience for the Helping Families Team Members at the NHS.

The objective of this project is to produce a mobile app that all members of the team can download. This would need specific features, for example being able to log appointment start and end times, check in and out of appointments, emergency functions allowing contact to be made with the entire team as well as sharing the users live location, reminders for the team to check in on clinician etc. When creating appointments, a buddy request will be sent and the buddy, when accepted, will be responsible for keeping in touch and ensuring safety. It should be an easy way for clinicians in the community to let their team know they are safe, and for the team to be able to act quickly if required.

Laravel, Apache, MySQL, PHP, HTML, CSS

Developed a web application using the LAMP stack and Laravel PHP framework for a Game Rentals society, enabling them to organize and rent products via the website.

While designing the website, consistency and usability where the two main criteria we considered. Each page on the website (other than the home page) is consistent in design. Each of them is laid out as a heading followed by a breadcrumb. The breadcrumbs help with accessibility and tell the users about their current location on the website.

Each card in the products page has been carefully designed. It makes sure that each card is consistent regardless of the name of the product, the description or even the image size.

Users can access all options simply from the navbar, except for the reports which can be found on their dashboard. All the necessary information is displayed always to users.

In addition to this, on completion of any action, or on performing an invalid action, the user is given alerts. Green (success) alerts for successful completion of a task, red (danger) alerts for attempting to execute an invalid task and yellow (warning) alerts in case of Exceptions in the code.

React Native (JavaScript), Android, iOS

Developed a React Native application for Capo scale conversions

Capo for Dummies is a tool which helps guitarists realise how Capo positioning will affect their music.

Capo is a device that shortens the length of all the strings at the same time, effectively changing the pitch. Depending on the positioning of the Capo, the pitch is changed differently.

This App calculates which key/chords you will be playing by placing the Capo in a particular position

Java

Created a text-based Role Playing Game (RPG) with multiple endings set in the future.

This is the Second Assignment for Programming Practice and Application (PPA). The game is a text based adventure game called ‘Lambda Corps’. It has various rooms which the player can go through, characters and items to interact with and objectives to complete. The game has been inspired from multiple sources. Some of the more important inspirations for the game are ‘The Stanley Parable’ and ‘Wolfenstein 3D’.

The game aims to create an interactive and interesting environment for the user. In addition to the many commands the user can use and the special features of the game, I have also incorporated numerous gaming and music references throughout the game.

The Game ‘Lambda Corps’ takes place in the future, on an Earth in which World War 3 has already taken place. World War 4 is going on, and the 2 sides, the United Federation and the Alliance, are using Cyber Warfare to fight. Due to this changed way of warfare, huge software companies have been added to the army. The protagonist’s name is William Joseph Blazkowicz, and he has just started working at one of these companies called Lambda Corps.

During the game, you can move through rooms, pick up or drop items and talk to different characters in the game. You are given multiple tasks throughout the game. The next task dialogue is only displayed once you have completed the current task. At any given point, if you forget the task, entering the ‘task’ command will tell you what the current task is. In case you get lost, then you can use the ‘help’ command to print the list of valid commands and the map of the facility, which also shows your current location.

As the game progresses, so do the days passed since you started playing. This gives the game a sense of progression. Certain tasks will make the game skip a few days, and so will going to the Recreational Room. However there is a limit to the number of days you can proceed to. Once the game reaches 75 days, you automatically lose the game and one of the possible outcomes is printed.

In the game you meet another employee in the facility, called Stanley. He tells you that the Chief of the Corps might be betraying the United Federation and leaking secrets to the Alliance. You are required to help him and perform the required tasks. You help him find proof of the Chief’s treason, or so he says. To stop the chief from leaking the code for the ‘Reboot Matrix’ of the United Federation Stanley has planted a bug in the Chief’s ‘Superluminal Transmitter’. Unfortunately because of his activities, Stanley has been banned from the facility. Now it is up to the player to activate the bug in the transmitter.

Once you reach the Chief’s office, you get to know that the Chief thinks that Stanley is the mole. The Chief tells you that Stanley has always been suspicious and now has turned out to be a spy for the Alliance who has already leaked information to them. Because of your interactions with him, suspicion has fallen on you as well. The Chief asks you to prove your loyalty by fixing the bug in the superluminal transmitter, so that he can send a message to the Headquarters. This is when the player is faced with a choice. You can either ‘activate’ the bug or you can ‘fix’ it. As it turns out Stanley was the mole along. When you fix the bug, you end up saving the United Federation and winning the game! If you choose to activate the bug, you end up losing the game. Other than this there are a few secret endings to the game. Once you have collected the CSB from the Server Room, you can actually go to the Chief and give it to him instead of giving it to Stanley. This will end the game as well and you will win since you were able to stop the attack. Also if you pick the rose found outside the facility and give it to Emily, then if you win the game, Emily asks you out on a date.

Java

Created a predetor/prey simulation of the Canadaian Boreal Forest with multiple terrains, animals, seasons, breding cycles etc, using Java.

For this assignment, we have chosen to simulate the Canadian ecosystem, closely resembling that of the Boreal forest. Animals like bears, wolves, elks, moose, wolverines and salmon populate the landscape along with 2 species of plants. Environmental factors like day and night, seasons, weather and terrain have also been simulated. The simulation contains a total of 3 different zones; the land, the shore and the lake. Bears, elks, moose, wolves and wolverines are land animals and are thus only allowed in the shore and the land. On the other hand Salmon are allowed in the lake and shore and cannot go to the land.

The animals can be carnivores, herbivores or omnivores. This determines what each animal will be able to eat. Each animal has a breeding season and if they are hibernating animals, then a hibernating season as well. The plants grow differently in different seasons and their growth is affected by the weather.

CREATED A GUI FOR A VISUALIZER FOR SUCH A MARKETPLACE DISPLAYING DATA GIVEN IN A CSV. USED JAVA, FXML, CSS AND GOOGLE MAPS API

This is an app created to explore available properties inside London using the property dataset of Airbnb. It’s designed to be easy to use, and powerful in the data it can display. Creating the piece of software used all the unique talents of Team Hummus, and we split the work evenly between us through GitHub and peer programming at Bush House.

The GUI carries a modern design with “matte” black backgrounds, bright pink elements and minimalism. Since the application is designed to be for Airbnb, the official “Airbnb pink” was used for text and borders to adhere to company design standards. The base idea of the GUI is to disguise the complexity of the backend behind a deceptively simple user interface.

MySQL

Created a database and performed various operations for the Voice London.

Implemented a relational model for a television talent show name "The Voice London", based on a previously created ER Diagram and set of requirements.

Performed different data manipulation and retrieval operations on the database.