Physicist working in machine learning. Los Angeles.
Four years in experimental high-energy physics, with ttHH searches at SCIPP and MIP Timing Detector work at UVA. Now building and probing neural networks. I'm interested in what models learn that nobody told them to learn, and in whether the uncertainty on a prediction means what it claims to mean.
Current work
- chess-transformer: legal-move generation is neither capacity-bound nor data-bound alone; each roughly doubles fully-legal games, both together take it from 4.4% to 51.8%. Linear probing the residual stream shows the 12-layer model's board representation peaks three layers before the output
- schrodinger-pinn: a PINN that discovers the ground-state energy as a trainable parameter rather than fitting to a known solution; staged Adam then L-BFGS recovers it to 0.0002% relative error on the infinite well
- gpt-from-scratch: epoch-boundary validation hid the loss minimum by 5,300 steps; step-interval evaluation cut val loss 1.63 → 1.46
- resume-tailor: role-specific résumés from a master document and a job posting; started as RAG, ended as full-context injection once retrieval proved wrong for a single-document corpus
Tools PyTorch · Python · C++ · ROOT/uproot · Docker
Publication Aromatic Copper Hydride Cages, AAAFM Energy 2020;1(1):16–26, first author
Elsewhere jotellechea.com
Open to research engineer and applied scientist roles.