Projects
Here are some projects I carried out for uni / fun.
Current research
- Turn-taking prediction in duplex models MSc dissertation work on whether sparse autoencoders can recover timing-related features in duplex speech-model representations.
- Circuit Atlas Prototype interface for exploring SAE feature clusters and their interactions across transformer layers.
- Moshi turn-taking data pipeline Supporting data pipeline for conversational audio: discovery, diarisation, transcription, Mimi encoding, and turn-yield labels.
Coursework and degree work
- Reward hacking under GRPO ATNLP coursework on GRPO for mathematical reasoning, focused on failure cases where the reward signal encourages shallow shortcuts.
- WFST-based ASR decoding ASR coursework implementing a WFST decoder with beam pruning, silence modelling, and KenLM language-model scoring.
- Histopathology image segmentation Computer vision coursework: UNet nucleus segmentation, then supervised and SimCLR-based classification.
- Biological knowledge graph construction Undergraduate dissertation using transformer-assisted triple extraction on biological abstracts, then turning the output into graph structure.
Other projects / experiments
- AI tutor with knowledge tracing A personal learning-tool prototype around concept graphs and learner state; still more design exploration than product.
- Emotion vectors for Gemma Exploratory interpretability pipeline for extracting emotion-related activation directions from Gemma activations.
- Obsidian and MCP automation infrastructure Local note-search and automation tooling around Obsidian, MCP, BetterTouchTool, and personal knowledge workflows.
- Snake AI Small Snake AI experiments, from older deep Q-learning code to newer checkpoint and evaluation work.
- Pathfinder Swift pathfinding and spatial-data-structure experiments, including quadtree and floorplan/SVG work.
- Speaker diarisation and claim verifier Rough speech tooling for diarising and transcribing long audio, then experimenting with claim extraction over the result.