Mostly: turning data into decisions. Currently finishing a research Master's and
tinkering with applied ML side-projects.
Climate downscaling with hierarchical denoising
Research Master's, Computer Science
Global climate models are coarse — typically 25–100 km per grid cell —
which is fine for global trends but useless if you need to know what's about to happen
over your specific catchment, vineyard, or grid asset. My research uses a hierarchical
denoising approach to take that coarse model output and produce high-resolution
predictions that match the underlying physics.
Precipitation (var 4) over training — the model's
prediction on the left, and the residual
against ground truth on the right. Watch the prediction develop fine-scale
structure as training progresses; the residual stays low-magnitude throughout.
Fish-GPT
Side project · vision + LLM
A toy that wires a vision model into a chat interface so you can point a camera at a
fish (or anything, really) and ask follow-up questions. Built mainly to see what the
failure modes of current vision-language models actually look like on natural,
non-curated input.
Short demo — identification + Q&A on a single fish image.