Scientific data analysis
- Error propagation & measurement uncertainty
- Analysis algorithms: reconstruction, classification, component decomposition & stress tests
- X-ray diffraction / scattering data
- Monte Carlo simulations
MEASUREMENT MODELS · ERROR PROPAGATION
Which measurement errors change the answer?
Data analysis through models of the physical processes behind your measurements.
RESEARCH SUBJECTS
Academic research
- Radiation chemistry & radical reactions in liquids
- Electron solvation, attachment & charge transfer
- Ultrafast processes: pulse radiolysis & transient spectroscopy
PhD in Chemistry · PhD in Physics
Senior Researcher · Institute of Chemical Physics
CNRS / Université Paris-Saclay
Academic research focuses on radiation chemistry, pulse radiolysis and time-resolved spectroscopy. Scientific R&D consulting includes an ongoing collaboration with EosDx on signal analysis, reconstruction and classification.
Profile on LinkedIn ↗How I can help
A physical model guides which correlations to test and what could explain them.
Stress-test your algorithms
Independent tests of when an algorithm works, when it fails, and which assumptions or measurement conditions explain the difference.
Analyse complex, unstructured data
Find candidate correlations; test them against physical mechanisms, measurement artefacts and dataset shift.
RESEARCH EXAMPLE
MATUR / EOSDXSignal representation matters.
X-ray scattering from ex vivo breast tissue: comparing signal representations for benign/cancer classification.
The 2025 study found better performance with 2D Fourier representations at wide angles; radial integration was comparable at small angles.
2025 co-authored study ↗Contribution, study limits & publications
The 2025 comparison covered radial and Fourier representations, preprocessing and classifiers across two scattering geometries. Repeated patterns from one patient are not independent patient cases.
The 2024 paper credits Sergey with methodology, software, data curation, visualization and preparation of the original draft. Author contributions ↗
A measurement model can test how noise, background and calibration affect the pipeline. This is a proposed next step, not an unpublished Matur result.
2025 · Benign/Cancer Diagnostics Based on X-Ray Diffraction: Comparison of Data Analytics Approaches ↗ 2024 · Vitacrystallography: Structural Biomarkers of Breast Cancer Obtained by X-ray Scattering ↗TRY A MODEL
Noise or calibration drift?
Noise broadens distributions. An offset shifts them. Different causes need different tests.
Open the interactive example
Compare
MEASURED FEATURE DISTRIBUTIONSFIG. 01
Class AClass BFixed threshold
Shaded tails: expected misclassification · feature units are arbitrary
Noise standard deviation σ.
Shift b applied to both classes.
Expected balanced error—
Class A —Class B —
Synthetic Gaussian model · fixed threshold · equal class weights. Not clinical or Matur data.
WHAT CHANGES
The reference case contains measurement noise but no calibration offset.
WHAT TO TEST
Repeated measurements can estimate random noise; a reference sample can test calibration.
Model assumptions & limits
True feature: z | A ∼ N(−0.8, 0.35²), z | B ∼ N(+0.8, 0.35²). Measurement: x = z + ε + b, where ε ∼ N(0, σ²) is independent of z. The rule assigns x < 0 to A and x ≥ 0 to B. Both classes have equal prior probability.
Displayed errors are Gaussian tail probabilities calculated analytically, not validation-set estimates. Baseline spread 0.35 is held fixed. The threshold is not retrained. Real spectra may have correlated noise, nonlinear responses, and population shift.
This is an illustrative model, not MATUR data or a clinical performance claim.
Another example: an uncertainty band
MODEL AND SIMULATED MEASUREMENTSFIG. 02
Known signal modelSimulated measurementCentral 95% measurement interval
The band needs a model.
Each point is drawn from y(q) = m(q) + ε, with independent ε ∼ N(0, 0.045²). The shaded band is m(q) ± 1.96 × 0.045.
It covers 95% of individual measurements at each q under this model. It is not a confidence interval for a fitted signal or a simultaneous band across the whole curve.
Synthetic illustration · arbitrary units · the signal model is known.
OTHER INTERESTS
Open data.
Prisons of the World
Prison locations with sources and coordinate-quality flags for research and human rights. Locations alone do not establish detention conditions.
Dataset on Kaggle ↗Scientific & medical accelerators
A partial registry with official-source evidence, distinguishing current, historical and planned facilities.
Registry on Kaggle ↗CONTACT
Let’s look at your problem.
Describe the measurement, question and current difficulty. Start with a non-confidential summary.
Paris, France · Research team · NDA available