Expertise & Services

Targeted Optimization

Profiling and optimizing scientific codes to reduce computation time without sacrificing accuracy. I also enjoy finding algorithmic tricks that reduce problem complexity.

Languages and Tools

• Fortran & OpenMP — For native, parallelized high-performance computing
• Python & Numba — My daily approach: writing Python code quickly with performance far exceeding standard Python thanks to JIT compilation
• C/C++ — Proficient, quick to pick back up depending on project needs
• CuPy, JAX, PyTorch — Currently learning for GPU acceleration

Experience

I worked on the computing cluster at the University of Franche-Comté (Besançon, France) for my simulations. My daily work involves optimizing optical and interferometric simulations as part of my PhD to be able to run millions of simulations on a laptop before having the time to make a coffee.

Hybrid Models

Combining physical knowledge with learning models to improve data analysis and prediction, primarily using MLP and CNN architectures, with a growing interest in Physics-Informed Neural Networks (PINNs).

Concrete Use Cases

• Trans-Neptunian object detection — Machine learning algorithm to assess our detection capability as a function of object albedo
• Protostellar cloud characterization — Neural network trained on numerically simulated images to recover temperature and pressure at the core of a cloud
• Interferometric calibration — ML approach to automatically determine the phases to inject into a photonic component to reach its optimal operating point
• Astronomical source detection — Attempted reproduction of a machine-learning-based S-Extractor

Approach

Selection of lightweight architectures, well-controlled datasets, cross-validation and explicit uncertainties to avoid overfitting.

Analysis and Models

Building robust data pipelines to explore, clean and model experimental or simulated datasets, regardless of data type: time series, images, spectra or signals.

Approach

• Solid statistics — Uncertainty estimation, hypothesis testing, metrics adapted to weak signals
• Interpretable models — Regularized regressions, anomaly detection, dimensionality reduction
• Reproducibility — Clean notebooks, version tracking and parameter traceability

Experience

This type of work is at the core of my PhD, particularly around statistical detection of faint signals in high-contrast interferometric data.

Tools

Scientific Python (NumPy/SciPy, pandas), clear visualization (Matplotlib, Seaborn), ML models when they bring measurable value.