TRACE: reward hacking in RL fine-tuning
A controlled study of reward hacking: SmolLM2-360M trained with SFT then GRPO against a weak and a strong verifier, with logit lens and probes to find where the two models diverge.
All Projects
Vocabscapes: multi-agent puzzle generation
A five-agent word-puzzle pipeline (writer, judge, improver, finalizer, explanation) with bounded retries and cached variants per level.
An LLM eval harness for puzzle generation
LLM-as-judge harness across four model pairs, calibrated against blind human ratings.
maybehere: group restaurant discovery
A live no-signup app for planning where a group eats: 97K+ place profiles across ~200 UK towns, semantic search on BGE-small, under 400ms per query on free-tier infrastructure.
Rosetta & OnePiece: portable LLM representations and memory
A fixed-size, model-agnostic embedding representation (0.955 MRR) and a ~1K-parameter memory scorer that matches a cross-encoder reranker with ~278,000× fewer parameters. Pre-prints in progress.
Social media recommendation engine
A TikTok-style recommender built from scratch: scoring, Gemini embeddings, FAISS indexing and the exploitation-exploration trade-off.
Privacy-preserving fraud detection with federated learning
Undergraduate dissertation: federated training across 3 clients on Ethereum transactions with 90%+ class imbalance, reaching 99.7% AUC-ROC.
Making Regression Make Sense: an Interactive Visualiser
An interactive tool for seeing regression: tweak slope, intercept, polynomial order and learning rate, and watch L1 and L2 fits update in real time.
UNDP climate survey analysis
Predicting and segmenting support for climate action across 73,000+ respondents in 73 countries: 89.76% AUC-ROC with a random forest, six country typologies with k-means.