🔬 Academic Research & Generative AI
🎨 Generative AI for Image Inpainting (MEng Thesis)
Context: MEng Research Project, Imperial College London. Download Thesis PDF:
Description: This Master’s thesis explored novel techniques for text-conditioned image manipulation by enhancing the object replacement and inpainting process using advanced segmentation and diffusion models.
- Core Architectures: The work involved exploring and benchmarking different Diffusion Model architectures to accurately fill masked regions of an image using contextual information.
- Novel Methodology: Developed and refined a novel methodology for improving GenAI image inpainting algorithms by integrating secondary AI mask generation.
- Pipeline & Achievement: Built a “find and replace” pipeline to segment objects using a text prompt, achieving state-of-the-art visual coherence for the inpainting process.
🧠 Masters Thesis (Future Focus)
Institution: Queen Mary University of London (Expected 2026)
Description: Currently specializing in Multimodal AI (Image, Audio, Language models) and Deep Learning Architectures, with a Predicted Classification of 1st (Distinction). The future project will focus on applied research within emerging Generative AI domains.
- Target Area: Aiming to undertake a project involving state-of-the-art Vision-Language Models (VLMs) or complex Computer Vision tasks.
- Core Disciplines: Deep Learning, Multimodal AI, and advanced Computer Vision architectures.
🛡️ Machine Learning & Optimization
📊 High-Impact Fraud Detection Model
Context: Personal Project (Kaggle Dataset)
Description: Developed and rigorously optimized a financial fraud detection system, focusing on maximizing the capture rate of high-cost fraud cases (minimizing False Negatives) in a severely imbalanced credit card fraud dataset.
- Model Architecture: Built and optimized an XGBoost classifier, chosen for its speed, interpretability, and robust performance on tabular data.
- Strategic Optimization: Optimized the model for the F-beta score ($\beta=20$)—a metric explicitly weighting Recall over Precision—to heavily penalize high-cost false negatives.
- Metric Achievement: Achieved an exceptional Recall of 95.7% (catching 44 out of 46 fraud cases) by strategically shifting the classification threshold to 0.0035. This result prioritized loss prevention over minimizing false alarms, a necessary trade-off that resulted in a Precision of 4.6%.
🗺️ Full-Waveform Inversion (FWI) for Geophysical Imaging
Context: Kaggle & Yale/UNC-CH ML Competition
Description: Engineered a sophisticated machine learning solution for a complex geophysical imaging problem, demonstrating adaptability to large-scale scientific datasets.
- Architecture & Techniques: Utilized deep learning (U-Net) and signal processing techniques to invert seismic waveform data.
- Data Handling: Showcased proficiency in handling non-standard, large-scale scientific datasets.
- Achievement: Achieved a highly competitive result in a specialized domain, demonstrating strong problem-solving and technical adaptation skills.
💻 Foundational Systems & Learning
📝 Foundational Deep Learning: GPT-2 Clone
Description: Implemented a foundational Generative AI architecture from scratch to gain a deep understanding of core transformer mechanics.
- Core Implementation: Built a decoder-only causal language model entirely in PyTorch, replicating the key components of the GPT-2 architecture (self-attention, residual connections).
- Data and Tokenization: Preprocessed the script of The Matrix to create the training corpus and utilized Byte Pair Encoding (BPE) for efficient tokenization.
- Objective: Demonstrated end-to-end knowledge of modern LLM pipelines, from data ingestion and tokenization to forward pass calculations and text generation sampling.
🥇 Kaggle Classification Competitions
Description: Showcase of proficiency in fundamental Deep Learning and ML techniques using widely recognized benchmark challenges.
- MNIST Digit Recognition: Achieved 99%+ accuracy on the MNIST classification benchmark using a custom-built Convolutional Neural Network (CNN), validating a strong grasp of foundational Computer Vision architectures and training stability.
- Additional Competitions: Participation in various other challenges (e.g., FWI) demonstrates adaptability across different data types (tabular, image, seismic signal) and modeling goals.