SmartCart v3

An augmented meal cart for in-hospital use that helps prevent malnutrition in hospital patients by monitoring their food intake. SmartCart v3 captures before and after images/weights of hospital meals, segments and classifies foods, estimates portions using depth and scale data, and computes calories, macronutrients and micronutrients using a recipe database. Note that the repo is private as the project is still ongoing.

Fire Extinguisher Allocation

A collection of RL policies developed for autonomous wildfire suppression using a fixed-wing drone. The agent learns to navigate a dynamic wildfire environment and optimally deploy fire extinguisher balls. This project explored the impact of different fire spread rates, observation spaces, reward functions and algorithms on agent performance. Note that the repo is private as the competition is still ongoing.

The Thief's Gamble

A decision-driven third-person adventure game where a player embarks on a medieval-themed quest, shaping their fate through choices and consequences.

The Polygon Dead

A third-person zombie shooter where a player battles waves of relentless zombies. Features multiple rooms, each introducing new challenges and enemy spawns.

Race Against Time

A fast-paced racing game where a player strives to beat their own lap record. Features both a single-player mode and a mode with AI-controlled opponents.

Ad Auction Dashboard

An analytics dashboard designed to evaluate and visualise the performance of advertising campaigns.

Tetrecs

A creative twist on Tetris featuring zero gravity and a multiplayer mode for a fresh strategic experience.

Orchestra Simulator

A simulator that recreates the sound of an orchestra performing a given composition, blending musical modelling and procedural audio generation.

Brain Tumor MRI Classification

A fine-tuned ResNet50 model trained to classify axial-view MRI scans of brain tumors. The model achieved an accuracy of 99% on the 4-class tumor classification task. This project explored the impact of data augmentation and transfer learning on model performance.

GTZAN Music Genre Classification

A collection of deep learning models trained to classify music genres using Mel Spectrograms. The best-performing model (a CNN with batch normalization) achieved an accuracy of 73% on the 10-class genre classification task. This project explored the impact of regularization techniques on model performance.

Text8 Word Clustering

A K-Means clustering model applied to the text8 dataset to group words based on their semantic relationships using a co-occurrence matrix. This project explored unsupervised learning, word representations and semantic structure discovery in text data.

Heart Disease Prediction

An XGBoost model trained to predict whether a patient has heart disease based on lifestyle and health indicators. Built as part of a hands-on workshop I delivered on building your first tree-based model, introducing participants to decision trees, ensemble learning and class imbalance.

Tesla Stock Price Prediction

A recurrent neural network trained to predict Tesla's stock price based on its value over the past 12 days. Built as part of a hands-on workshop I delivered on building your first RNN, introducing participants to sequence modelling, temporal dependencies and time series forecasting.

Simpsons Image Classification

A convolutional neural network trained to classify images of characters from 'The Simpsons'. Built as part of a hands-on workshop I delivered on building your first CNN, introducing participants to image classification, feature extraction and convolutional layers.

Titanic Passenger Survival Prediction

A neural network trained to predict passenger survival on the Titanic using features such as passenger class, sex and family aboard. Built as part of a hands-on workshop I delivered on building your first neural network, demonstrating the fundamentals of data preprocessing, model training and model evaluation.