AI PhD candidate at the Australian National University (expected 2027), working on generative, multimodal, and evaluation systems for decision-making under uncertainty. My research builds generative world models and spatial reasoning for perception and planning, alongside methods for evaluating what learned systems actually get right. My thesis, Into the Unknown, develops generative posteriors for reasoning about spatio-semantic uncertainty, advised by Rahul Shome, Dylan Campbell, and Stephen Gould.
Research
- Believing is Seeing: Unobserved Object Detection using Generative Models — CVPR 2025 · project
- FlatLands: Generative Floormap Completion from a Single Egocentric View — ECCV 2026
- MatterDoor: Sampling Zero-shot Spatio-semantic Priors using Generative Models — under review; arXiv:2510.11014; presented at the RSS 2026 FM4RoboPlan workshop
Open-source software
- TorchKAN — PyTorch Kolmogorov–Arnold Networks (spline, Legendre, Chebyshev, and convolutional variants) with post-training quantisation, CUDA execution, and Integrated-Gradients interpretability (190+ stars).
- Mangroves — Python package for depth-based hierarchical data management, type-aware ingestion, and CPU–GPU tensor movement.
- TuringTree — vectorless, on-device RAG (Ollama + Qwen) with 0–100 confidence scoring and abstention; co-developed by a four-person team during the Microsoft Global Intern Hackathon.
Applied & quantitative research — LLM-as-judge and RAG evaluation at Microsoft; Korean retail-flow modelling and market-microstructure ML at Optiver; financial-document retrieval and question answering at JPMorgan Chase. This quantitative work — probability, optimisation, temporal validation, and noisy-data inference — underpins how I build and evaluate AI systems.
Website · Google Scholar · LinkedIn · Resume