Results

The same discipline, applied to your business

Before building AI systems for companies, I measured and optimized scientific simulations until they ran 678.9 times faster. It is the same way I work on any process that becomes slow or expensive.

678.9x

Molecular transport on GPU

I started from a simulation of molecules moving through a narrow channel that already existed as a serial and an MPI version (by a colleague). With 4 and 8 cores, MPI sped it up 3.6x and 6.4x, respectively.
I ported it to GPU with CUDA: 678.9 times faster than the serial version.
That speed made it possible to sweep more than 14,000 configurations and fix a bias in lateral movement that could not be seen before.
It is the same method I apply to a client's system: measure where the time goes, redesign what slows it down and prove the improvement with numbers.

678.9x chart ↗CUDA code ↗

Bar chart of speedup versus the serial version: serial 1.0x, MPI with 4 cores 3.6x, MPI with 8 cores 6.4x and CUDA 678.9x — How much faster each version runs compared with the serial one. CUDA reaches 678.9x; MPI reaches 3.6x and 6.4x.

Scenario 1

Scenario 2

Scenario 3

Scenario 4

Four of the more than 14,000 scenarios I tested, each with different rules for entering and leaving the channel.

< 8 days

Tumor immunosurveillance

Peer-reviewed · IEEE ROPEC 2025

I modeled how the immune system interacts with a tumor: tumor cells, macrophages, T cells and blood vessels, each with its own rules.
I optimized how the simulation uses memory and parallelized it with OpenMP: 6 times faster across more than 7,600 runs.
The full analysis dropped from more than 32 days of computing to under 8, and the work was peer-reviewed and published in IEEE Xplore.

Read the IEEE paper ↗INAOE code ↗TME diagram ↗Academic poster ↗

Diagram of the biological tumor immunosurveillance model, designed by hand — The biological model: a tumor with a hypoxic core surrounded by M0, M1 and M2 macrophages, active, exhausted and naive T cells, fibroblasts, extracellular matrix, and blood and lymphatic vessels.

The simulation running: each dot is a cell that moves, interacts and responds to the tumor.

6 hours

Brownian dynamics (bachelor's thesis)

I simulated how positive and negative ions move around a charged particle, physics that requires computing a huge number of interactions at every step.
I wrote the simulation to split that work across hundreds of GPU cores with OpenACC.
What took 30 days in a normal run dropped to 6 hours. In the chart, the GPU runs 100 to 150 times faster than serial at every size.

GPU vs CPU chart ↗Trajectory ↗Thesis code ↗

Electrolyte simulation: positive and negative ions around a charged particle — A charged particle (white) surrounded by positive (red) and negative (blue) ions moving at random.

Chart comparing run times on serial CPU, OpenMP and GPU (OpenACC) — Run time by problem size (logarithmic scale): serial, OpenMP on CPU and OpenACC on GPU.

Tools and repositories

Open source and utilities

Production engineeringDemo on Hugging Face

infartos-mlops

End-to-end MLOps with a live demo, CI/CD and 35 tests.

View repository on GitHub

Automation and dataAutomatic extraction

docpipe

Extracts information from documents, ready to be used as a tool by an AI assistant.

View repository on GitHub

AI productDesktop tool

pdf_to_excel

Extracts tables from scanned documents into Excel with help from Gemini.

View repository on GitHub