Real-time conversational NPCs for VR
End-to-end AI and voice pipeline for talking NPCs in immersive VR: adaptive voice activity detection, LLM dialogue orchestration and procedural quest generation, at under 500 ms latency.
Projects
Open-source libraries and applied AI and measurement work. Code is on GitHub.lens-model: differentiable camera lens model
A compact PyTorch forward model of a real lens (distortion, field-dependent wavefront, vignetting, radiometry) fitted to images by photometric bundle adjustment, and benchmarked against the classical OpenCV pipeline.
nodes: run heavy jobs on your own PCs
A small command-line tool that turns a handful of personal GPU machines into a job queue: sync a project, queue a job, follow logs, pull results, all over Tailscale, SSH and rsync.
hpcom: GPU-accelerated optical fibre link simulator
Open-source Python library for fast, reproducible simulation of coherent optical communication links. Runs the split-step Fourier and Manakov solvers on TensorFlow/GPU, with speed-ups of up to 2000×.
nft-processing: nonlinear Fourier transform toolkit
Python toolkit for forward and inverse nonlinear Fourier transforms, digital back-propagation and windowed processing of long signals, built on a modified FNFT C library.
optcom-trees: forest-based nonlinear equalisers
Random forest, gradient boosting and XGBoost models that learn to undo the nonlinear distortion of coherent optical signals from a symbol’s neighbours, as a fast, FPGA-friendly alternative to neural-network equalisers.
triplet-fitting: Gaussian mixture models of nonlinear distortion
Python package that fits Gaussian mixture models to triplets of transmitted and received symbols from simulated optical links, across launch powers and many runs, to characterise nonlinear signal behaviour.