Problem: Rocket launches are extremely costly. Predicting the likelihood of successful landings allows SpaceX to plan reusability strategies and reduce mission costs.

Approach: Collected launch data from SpaceX API and web-scraping sources; cleaned and structured datasets; performed exploratory analysis with interactive maps and visualisations; engineered features (payload mass, orbit, flight site, booster type); trained classification models (logistic regression, decision tree, random forest, SVM); and evaluated performance with accuracy, F1, and cross-validation.

Results: Achieved strong predictive performance on test data (≈85–90% accuracy depending on the model). Produced clear dashboards and visuals of launch outcomes by location and year. Demonstrated how data science can inform aerospace engineering decisions and cost-saving strategies.

Problem: Wet-lab experimentation to optimise growth conditions is costly and time-consuming.

Approach: Discretised a continuum model for numerical computation; implemented simulations in Python & Julia; preprocessed large multi-modal clinical datasets; performed parameter identifiability analysis and evaluated statistical fit before generating forward simulations across environments.

Results: Delivered a reproducible simulation framework that forecasts tissue growth under varying conditions, reducing reliance on extensive wet-lab iterations.

Problem: Need to predict tumour-volume response under different dosing schedules, including combinations, from sensitive preclinical data.

Approach: Collaborated with GSK; adapted oncology PK/PD models from literature to capture combination effects; calibrated parameters to preclinical datasets using MATLAB, Python & Monolix; validated with held-out data.

Results: Produced regimen recommendations from fitted models and scenario analyses to inform experimental planning.

Problem: Selecting an appropriate deep‑learning architecture and data‑loading strategy is critical for real‑world image‑classification tasks. Understanding trade‑offs between frameworks (PyTorch vs Keras) and models (CNNs vs ViTs) is necessary before tackling applied problems such as waste sorting.

Approach: Implemented convolutional neural networks and vision transformers in both Keras and PyTorch, comparing training performance and efficiency. Built a PyTorch classifier from scratch and explored hybrid CNN‑ViT architectures. Conducted experiments on breast‑cancer classification and compared memory‑based vs generator‑based data pipeline. For the capstone, fine‑tuned pre‑trained models on a waste‑product image dataset, applying cross‑validation and hyper‑parameter tuning.

Results: Demonstrated how CNNs and ViTs perform across tasks, highlighting efficiency and accuracy trade‑offs. Achieved strong classification accuracy on benchmark datasets and improved waste‑classification performance via transfer learning. Provided insights into data‑pipeline choices and model selection for future projects.

Problem: A solid grasp of linear algebra, optimisation and probability is essential for designing and understanding machine‑learning algorithms. Many learners struggle to bridge theoretical concepts and practical implementation.

Approach: Worked through exercises inspired by Imperial College London’s Mathematics for Machine Learning series. Topics include vector spaces, orthogonality, projections, inner products and back‑propagation mathematics. Implemented algorithms such as K‑Nearest Neighbors and PageRank; practised neural‑network fundamentals via gradient‑based optimisation and fitting helical distributions. Additional exercises explore transformation matrices and reflections.

Results: Produced a set of well‑annotated notebooks that translate mathematical theory into code, serving both as a learning resource and as a reference for future projects. Completing these exercises deepened understanding of the maths underlying machine learning.

Problem: Restaurants require reliable APIs to manage reservations, menu items and customer data. Building such an API involves creating secure endpoints, handling CRUD operations and supporting authentication.

Approach: Built a RESTful API using Django and the Django REST Framework. Implemented serializers, token‑based authentication and CRUD functionality for reservations, menu items and customer data. Validated the API with Postman and deployed it locally as well as on GitHub.

Results: Delivered a functional back‑end that allows clients to create, read, update and delete restaurant data while ensuring secure access control. Successful Postman tests confirmed endpoint reliability, making the API suitable for integration with front‑end or mobile applications.

Problem: Investors and analysts often need to dissect historical stock data to understand trends, assess normality assumptions and develop predictive models. Such analyses must handle outliers, non‑normal distributions and provide interpretable metrics.

Approach: Conducted in‑depth exploratory data analysis on Apple (AAPL) price data, including summary statistics, missing‑value analysis, kurtosis assessment and correlation heatmaps. Applied normality tests (Shapiro‑Wilk, D’Agostino’s K², Anderson–Darling) and visualised the data through histograms, scatter plots, line plots, boxplots and candlestick charts. Detected and removed outliers, then fit a linear regression model using ordinary least squares and evaluated it with metrics such as MSE, MAE, MAPE and R².

Results: Generated comprehensive visual and statistical insights into AAPL’s historical behaviour, highlighting distributional properties and relationships between variables. The regression model provided a baseline predictive framework; metrics indicated reasonable fit while emphasising the limitations of simple linear models. The notebook serves as an educational reference rather than a trading recommendation.

Problem: EEG signals are noisy and non-stationary, making reliable intent decoding difficult.

Approach: Built a MATLAB pipeline with denoising and feature extraction; trained classification models with train/validation/test splits; evaluated with accuracy, sensitivity and specificity.

Results: Achieved robust decoding performance with clear metric reporting, supporting feasibility for assistive interfaces.

Problem: Need accurate, interpretable segmentation for downstream cardiac analysis.

Approach: Prototyped multiple learners in RStudio (logistic regression, random forest, clustering); compared pipelines via cross-validation; reported per-class metrics.

Results: Selected a high-performing segmentation approach suitable for reproducible clinical analysis workflows.