Research
Finding Attack Patterns in Jailbreaking of Large Language Models
Jul 2024 – Oct 2025Studied recurring patterns in successful prompt injection attacks against an LLM, using the Tensor Trust dataset of real player-submitted attacks. Embedded and clustered 7,178 unique successful attacks (sentence embeddings, t-SNE, HDBSCAN), finding 118 semantic clusters; the most repeated attacks ranged from direct instruction injection to rare-token and repeated-character spam.
Fine-tuned Matcha-TTS (an encoder-decoder TTS model trained with Optimal-Transport Conditional Flow Matching) on a single target speaker's voice to let laryngectomy patients regain their voice identity. The 14M-parameter model reached a Mean Opinion Score of 3.66/5 and a real-time factor of 0.038 — fast enough for live conversation — then was exported to ONNX for lightweight, on-device inference.
Leveraging Deep Learning for Tracking Calories in Food
Oct 2022 – Mar 2023Trained a deep-learning model (MobileNetV2 transfer learning, 56 food classes, ~91% test accuracy) to recognize food from a photo and estimate its calorie content, aimed at replacing error-prone manual calorie tracking. Built as a Streamlit app with automatic calorie lookup.
Utilising Machine Learning to Detect Debris in Lakes and Rivers
Sep 2020 – Feb 2021Flew a drone over local water bodies (including Cypress Creek) and ran two pretrained object detectors, Darknet and YOLOv3 (no fine-tuning), to identify debris, then manually checked detections against what was actually in each scene. YOLOv3 reached ~32% accuracy vs. ~7% for Darknet; common materials like glass, plastic, and cans were identified far more reliably than irregular debris. Also tested robustness against Gaussian noise and partial/cropped images.