Portfolio

Data Science & Machine Learning

Other

EIFO Data Extraction/Web Scraping

Built a Python script to extract country risk and cover polcies from EIFO website.
Used BeautifulSoup and Selenium (for dynamic website features) for data selection and extraction.
Cleaned and transformed data in Pandas.
Exported data to Excel output sheet.

Created a Convolutional Neural Network in TensorFlow to predict the class of a handwritten character.
Artificially constructured characters with a strikethrough.
Used measures to avoid overfitting a Computer Vison task, such as Dropout, Batch Normalization, Kernal Regulization, and Data Augmentation.
Achieved a 99% prediction accuracy on the test data.
Model deployed for real world application of Education Scoresheet recogniton.

Created a Neural Network Stacking Classifier using Scikeras KerasClassifier and Scikit-learn’s StackingClassifier.
Used analysis of feature outliers and distrubtion to improve feature engineering.
Combined Catboost, XGBoost, LightGBM and Random Forest Classifier models to create a balanced ensemble model.
Utilized Tensorflow 2.0 Sequential Neural Network for improved ensemble weights.
Achieved a final score of 0.87583 (compared with best of 0.87946), placing 283 out of 1910 submisions.

Created a Fine-tuned task specific model using Llama3 and Unsloth to create Dungeons & Dragons statblocks and game elements.
Deployed on Hugging Face for accesible Training Dataset and LoRA Fine-tuned Adaptors.

Explored forecasting daily kWh consumption for a workplace EV charging program and implemented naive, neural network, and LightGBM prediction models.

Bitcoin Price Forecasting

Implemented time series models (Conv1D, Bidirectional LSTM, GRU) for Bitcoin price forecasting, using metrics (Loss, MAE, MSE) to assess performance.

Structured Medical Text Classification with NLP

Developed a multi-class NLP model utilizing Universal Sentence Encoder, Conv1D character embedding, and text position features to classify medical literature sentences, enhancing readability with structured paragraphs based on classification headings.

Transfer Learning in Food Image Classification

Developed computer vision model on Food101 dataset utilizing EfficientNetV2 transfer learning to achieve over 80% accuracy in categorizing 101 types of food images.

Utilized Machine Learning techniques with Scikit-learn to predict character races based on ability scores and features from a Dungeons & Dragons dataset.
Explored various models for selection, Random Forest, Logistic Regression, Support Vector Classifier (SVC), K-Nearest Neighbors (KNN), and AdaBoost Classifier.
Conducted in-depth feature analysis and model evaluation.
Fine-tuned hyperparameters of the best performing classfier through manual tuning, RandomizedSearchCV and GridSearchCV.
Achieved an accuracy of 67.25% in predicting character races, improving on 51.8% from the initial dataset.

A financial assessment and report for ‘DnD with Toph’, an online Dungeons & Dragons adventure service, using Python and Pandas to clean, transform, and analyze the data.
Used data to observe growth and profit trends to optimise future scheduling and projects.
Skills include Data Manipulation & Analysis using Python & Pandas and Data Visualization with Matplotlib.

Developed a Python-based Zombie Survival Game, the mechanics involve managing resources, mission outcomes, and end-of-day round Zombie attacks.
Object-oriented programming in Python to create Class Models for Camp, Locations, and Players.
Utilized smaller functions for specific game mechanics and larger functions to implement and run game flow, mission outcomes, and end-of-round events.
Employed random library and computer decision logic trees for computer-controlled players.