Research
-
2026
Lynx: Progressive Speculative Quantization for Accelerating KV Transfer in Long-Context Inference
-
2025
Lightweight Online Adaption for Time Series Foundation Model Forecasts
-
2025
Higher Acceptance Rates for Speculative Decoding with Randomised Drafting
-
2025
Performance of Zero-Shot Time Series Foundation Models on Cloud Data
-
2025
Robust Loss Functions for Machine Learning in the Presence of Noisy Labels
- 2024
- 2024
- 2023
The full list, with citations, is on Google Scholar.
Side projects
-
Oct 2026
A record snake in the 13-dimensional cube
A snake-in-the-box of 2946 edges in thirteen dimensions, improving on the previous best of 2938. Watch it draw itself, or verify it in your browser.
-
Sep 2026
A spaced-repetition web app for learning the meanings of all 2,136 jōyō kanji, each taught after the parts it's built from, with mnemonics. I use it on my iPhone.
-
Apr 2026
Numbers k that divide the reversed binary concatenation of 1 to k. The first three terms, 1, 5 and 337, had stood alone; a parallel GPU search found the fourth, 58088321, and showed there are no others below 108.
-
Earlier
Designing an AI that solves Wordle in as few guesses as possible.
Background
- NowMachine learning researcher, Sakana AI, Tokyo
- 2024–26Researcher, Huawei efficient LLM inference
- PhDUniversity of Edinburgh robust learning with noisy labels
- 2022BlackRock AI Labs time series modelling
- MMathMathematics, University of Warwick with a year at LMU Munich