Tomás Erdmannsdörffer
CV

I teach physics to neural networks.

I'm Tomás, a machine learning engineer in Santiago, Chile. I build physics-informed neural networks for the mechanics of heart tissue, and compressed versions of them that still get the answer right. MSc student at PUC Chile; visiting researcher at the University of Graz until August 2026, still working with the group from here.

See the five demos Read the CV

Fig. 1 The average adult heart of Rodero et al. (2021), a four-chamber simulation mesh built from CT scans, decimated for the browser. Roughly the geometry my networks learn to deform.

Five demos that run in your browser.

No backend, nothing pre-rendered. Each page is a solver or a model running on your own machine while you watch. The only thing that leaves the page is the answering step of the RAG demo, and that uses a key you bring.

PINN Playground

Train a physics-informed network on the 1D Burgers equation, live. There's no dataset: the loss is the PDE residual, computed with automatic differentiation on your GPU. Give it two minutes and a shock forms.

TensorFlow.js, WebGL, automatic differentiation

Open the demo
∂u/∂t + u·∂u/∂x = ν·∂²u/∂x²

Ask Research

Retrieval-augmented answers over a corpus of 24 scientific ML papers. BM25 retrieval runs in the page; plug in any OpenAI-compatible key (Groq's free tier works) and a model answers from the passages it found.

BM25, retrieval-augmented generation

Open the demo
query → retrieve → answer

FNO vs. Solver

A neural operator and a finite-difference solver get the same Burgers problem. The solver marches through every time step; the operator answers in one call. The speed-up on screen is measured on your machine, not quoted from a paper.

Neural operator, Cole-Hopf transform

Open the demo
solver, every step operator, one call

Fluid Playground

Jos Stam's stable-fluids solver in plain JavaScript. Drag to stir ink through a divergence-free velocity field that gets re-solved every frame.

Navier-Stokes, canvas 2D

Open the demo
∇·u = 0

Optimizer Race

SGD, momentum, RMSProp and Adam dropped onto the same 3D loss landscape. Click the surface to move the starting point and see which ones get stuck.

Adam, three.js

Open the demo
θ ← θ − η·∇L

Scientific computing first, machine learning second.

Tomás, in a cap and round glasses Tomás, Graz, winter 2026

I got into ML through scientific computing, and most of what I've done since has been about making neural networks respect the physics they're approximating.

My undergraduate thesis at Universidad de los Andes compressed physics-informed networks (distillation, pruning, quantization) for non-Newtonian fluid simulation; the manuscript is in preparation. Since 2025 I've been doing an MSc at PUC Chile with Francisco Sahli, on PINNs for cardiac fiber mechanics. That work took me to the University of Graz for the first half of 2026, where I worked with Federica Caforio's group on PINN architectures for cardiac mechanics in JAX. I'm back in Santiago now and the collaboration continues remotely.

The rest of the time I build things to find out whether they work: the demos above, a series of language models under 16 MB, a file-transfer system that's still running at a mining company.

Based
Santiago, Chile
Languages
Spanish (native), English (C2), German (A1)
Status
Open to ML/AI roles, remote or on-site

Right now

  • ResearchCardiac motion estimation with PINNs in JAX, with the Graz group.
  • ExploringLanguage models under 16 MB: custom tokenizers, quantization.

Where I've been.

  1. Feb – Aug 2026

    Visiting researcher

    University of Graz, Dr. Federica Caforio's group

    PINN architectures for high-dimensional PDEs in cardiac fiber mechanics, in JAX: differentiable PDE simulation and multi-view reconstruction scored with MSE and SSIM. We still work together remotely.

    JAX, PINNs, finite elasticity, MRI
  2. 2025 – present

    MSc in Engineering Sciences

    Pontificia Universidad Católica de Chile, advised by Prof. Francisco Sahli

    PINNs for cardiac fiber modeling. Grew out of the undergraduate thesis on compressing them.

    JAX, PyTorch, PDEs, biomechanics
  3. Jan – Feb 2025

    DevOps intern

    Citi, Chile Tech Center

    Kept three legacy applications running and reworked the production CI/CD pipeline so deployments took less time and fewer steps.

    Jenkins, Bitbucket, Docker
  4. 2024 – 2025

    Undergraduate thesis

    Universidad de los Andes, manuscript in preparation

    Knowledge distillation, pruning and quantization applied to PINN surrogates of Carbopol (non-Newtonian) flow. Built the baseline PINN, then measured what each technique costs in accuracy.

    PyTorch, PINNs, quantization-aware training, CFD
  5. Jun 2023 – Dec 2024

    Teaching assistant, six courses

    Universidad de los Andes, CS department

    Operating systems, low-level programming in C/C++, mobile apps, web technologies, automata and computability, programming fundamentals. About thirty students a semester.

    C/C++, Python, Linux
  6. Dec 2023 – Feb 2024

    AI research intern

    Falabella Retail

    Prototyped automated product descriptions with LLMs, replacing a manual process. The proposal kept being developed after I left.

    LLMs, Python, prompt design
  7. 2020 – 2025

    BSc Civil Engineering in Computer Science

    Universidad de los Andes, minor in innovation

    Coursework in AI, LLMs, computer vision, algorithms and competitive programming, databases and web technologies.

    Python, C/C++, PyTorch, SQL

Beyond the demos.

  1. 2025 – present

    Cardiac motion with PINNs

    Lab work

    Physics-informed networks for cardiac biomechanics, built on jaxpi. The research behind the heart at the top of this page.

    JAX, PINNs, jaxpi
  2. 2025 – present

    Tiny language models

    Exploration

    Transformers under 16 MB: custom tokenizers and quantized weights, scored by bits-per-byte on FineWeb. Trained at home on an RTX 5060, which meant living on PyTorch nightlies for a while.

    PyTorch, CUDA, quantization, tokenizers
  3. Sep 2024

    Turbo Files

    In production at a mining company

    Fault-tolerant transfer of large files in Django: files travel as fragments so an intermittent connection doesn't kill the transfer. Google Drive and S3 as backends.

    Django, AWS S3, Drive API

What I work with.

Day to day: Python, PyTorch, JAX and PINNs, with CUDA when it matters. HuggingFace and TensorFlow when a project already lives there; RAG pipelines and agents; quantization and distillation for making models small.

Also fluent in: JavaScript, C and C++, C#, SQL. Django, Node.js, React, REST APIs, Ruby on Rails. Docker, Git, AWS (S3, Lambda), Jenkins, CI/CD, Linux and WSL2.

Write to me.

For roles, collaborations, or questions about any of the demos. Part-time while I finish the MSc, full-time after; remote or on-site anywhere.

GitHub LinkedIn CV