SKILLS
My Expertise
oscarfgavin@gmail.com
PROJECTS
My Work
Landmark Localisation
Drawing on the prominent UNet framework as a source of inspiration, I adapted the encoding layer to incorporate Transformers, surpassing the use of solely convolutional layers. By doing so, the model's ability to comprehend the spatial associations among landmarks improved, leading to enhanced precision in their localization. The model was then trained on a publicly accessible biomedical dataset, and ultimately produced a noteworthy average Intersection over Union (IoU) score of 0.92 during testing.
Cardology.io
Cardology is a unique platform that leverages cutting-edge technology to create customized flashcards from any PDF document. Our proprietary system uses GPT-3 to extract the most important information from your document and transform it into a set of flashcards for easy studying. With our intuitive user interface and advanced spaced repetition algorithm, you can easily revise and retain the information you need for your studies or work.
Chess Piece and Board Position Classifier
For the chess piece classification project, I had the task of accurately classifying chess pieces with varying levels of noise and board positioning. I utilized advanced techniques such as Principal Component Analysis (PCA), divergence, and K-means clustering to accomplish this task. Following a thorough testing and analysis process, the model achieved a 98% accuracy rate for chess piece classification and 90% accuracy for board positioning.
Sentiment Analyser for Rotten Tomatoes Reviews
For my sentiment classification project on Rotten Tomato reviews, I was tasked with classifying reviews on a scale ranging from 0 (negative) to 4 (positive), with 2 indicating neutral sentiment. To accomplish this, I utilized a Naive Bayes classifier, as well as advanced techniques such as Term Frequency-Inverse Document Frequency (TF-IDF) and Jelinek Mercer smoothing. What sets my approach apart is that I hard-coded all techniques, without relying on libraries such as scikit. Through extensive testing and analysis, I was able to achieve a record-breaking accuracy rate of 75%, surpassing the results achieved by my peers. I attribute this success to my innovative approach, which involved finding words most likely to produce correct classifications and weighting them based on their information content.
