👋 hi, I'm Tomás
visiting researcher @ uni graz
msc eng. @ puc chile
I teach physics to neural networks.
I train, compress, and ship ML models. Everything 3D on this page runs live in your browser, go poke it.
I train, compress, and ship ML models. Everything 3D on this page runs live in your browser, go poke it.
I'm a few months into a visiting research stay at the University of Graz, working on physics-informed neural networks for myocardial tissue modeling.
Beyond research, I like exploring what programming can do. Web experiments like the five live demos on this page, neural-network architectures, backends, whatever looks fun to build next.
Train a physics-informed neural network in your browser to solve the 1D viscous Burgers equation. No labeled data: the network learns from the PDE itself, via automatic differentiation, on your GPU. Watch the loss fall and a shock wave form.
RAG system over a curated corpus of 24 scientific ML papers. Retrieval runs in your browser; bring any OpenAI-compatible key for the answers.
Watch a neural operator solve Burgers 10-50× faster than a finite-difference solver. Same answer, fundamentally different algorithm.
A real-time Navier-Stokes solver (Stam's stable fluids) in pure JavaScript. Drag to stir candy-colored ink through a divergence-free velocity field.
SGD, Momentum, RMSProp & Adam race down a 3D loss landscape. Click the surface to drop them somewhere new. Who escapes the local minima?
Designing PINN architectures to solve high-dimensional PDEs in cardiac fiber mechanics using JAX.
Researching PINNs for cardiac fiber modeling applied to PDEs, the foundation of my work in Graz. The MSc deepens what started as a thesis on compressing PINNs.
Maintained and optimized 3 legacy applications, improved the production CI/CD pipeline to reduce operational complexity and deployment time, and collaborated on code migration via Bitbucket.
Applied knowledge distillation, pruning, and quantization to PINN surrogates for non-Newtonian (Carbopol) fluid simulation. Built the baseline PINN end-to-end, then quantified precision vs. efficiency trade-offs across techniques.
Supported ~30 students per semester across Operating Systems, Low-Level Programming (C/C++), Mobile Apps, Web Technologies, Automata & Computability, and Programming Fundamentals.
Researched generative AI tools for retail. Built a functional prototype for automated product descriptions, turning a manual process into an automated flow. Delivered a technical proposal that continued in development after the internship.
Coursework: Low-Level Programming, AI, LLM, Computer Vision, Algorithms & Competitive Programming, Databases, Web Technologies, Data Structures.
Lab research applying physics-informed neural networks to cardiac biomechanics.
Designing language models under 16MB: custom tokenizers and quantized Transformer architectures, benchmarked by bits-per-byte on FineWeb.
Fault-tolerant data transfer system shipped to a mining company in production. Fragments large files for resilience over intermittent connectivity.
A connected parcel locker we built as a team: when a package arrives the locker is opened remotely from the cloud, and it only unlocks when the right code is entered on the web app.
Let's build something that actually runs.
Open to ML/AI engineer roles, remote or on-site anywhere. Part-time now, full-time after the MSc.