Cage-based mesh deformation
Real-time mesh deformation driven by a coarse cage, with cage handles that move together smoothly.
CodeProjects
Research
Batched VisACD
A batched CUDA and OptiX rewrite of VisACD for high-throughput, deterministic convex decomposition of many meshes on one GPU, with asynchronous per-mesh pipelines and fused GPU plane search.
CodeVisACD paperParameterising context clues
An auxiliary network turns context and chain-of-thought into LoRA parameters at inference time, to speed up context-driven LLM inference.
CodeLearnable activations in physics-informed networks
Extends PINNsFormer's activation ablation to learnable activation functions for PDE learning.
CodeInteresting-state learning for Mario
An RL agent that replays states where it performs badly as targeted practice, built on a custom NES emulator.
Also
- Distribution shifts in self-supervised ViTs with test-time fine-tuning
- Are sequential models fantastic simulators?
- Can self-supervised models unlearn themselves?
- Downstream feature masking for contrastive models
- GAN dynamics via Kolmogorov–Arnold networks
- Real-time script generation with LLMs fine-tuned on small datasets
- Real-time audio diffusion via freeze maps
- Hidden correlations in Sequence Read Archive bioproject metadata
- Automatic labelling of data clusters
- Syntactic dependency trees for semantic equivalence
- Recurrent threshold neural networks
- Video games for early-years maths
- PPO in procedurally generated environments
Engineering
Mini-C compiler
A compiler for a subset of C with constant folding and propagation, loop unrolling and mem2reg passes.
Flipwise
A cross-platform app for planning and splitting shared expenses, built by a team of seven.
CodeAlso
- OS internals: user-level threads, virtual memory paging and a FUSE file system
- Type inference for a Haskell subset
- Part-of-speech tagging with HMMs and Viterbi
- Battleship solitaire solver with AC-3
- Breakout in MIPS assembly