Portfolio

My projects

A selection of work in computer vision, machine learning, data engineering and software development — from an R&D internship at Siemens Healthineers to academic and freelance projects.

Flagship project

Ovarian Follicle Segmentation in 3D MRI

Siemens Healthineers · R&D internship, Feb – Aug 2024

Detected ovarian follicles labeled and outlined on an MRI slice
PythonYOLOv8TensorFlow / Keras / PyTorchOpenCVMONAINiBabelSimpleITKDICOM / MRI
Problem
Fertility evaluation relies on manually counting and measuring ovarian follicles on MRI — a slow, repetitive, operator-dependent process.
Approach
An autonomous, end-to-end project: prepared 3D MRI data (normalization, interpolation, resampling, filtering, augmentation), evaluated 2D/3D U-Net architectures, then built a YOLOv8-based instance-segmentation pipeline trained and evaluated on data from multiple patients. Developed post-processing methods to compute follicle dimensions, volume, location and count, with a rule for flagging potential cysts based on diameter. Built a Python interface to visualize results and export CSV reports and annotated videos, working closely with clinicians throughout.
Result
~90% precision and 86% recall on segmentation masks and bounding boxes on the evaluation dataset. Analysis time dropped from roughly 10 minutes manual to ~10 seconds automated.
See full experience details →
Brain Mapping Research Center · Neuroimaging & ML

Brain Age Prediction from Structural MRI

Research project at the Brain Mapping Research Center — an honest, iterative pipeline for predicting brain age from MRI-derived features.

PythonScikit-learnFastSurferPandas / NumPyCross-Validation
Problem
Predicting "brain age" from MRI is a well-known biomarker for neurological and cognitive health, but naive pipelines are prone to regression-to-the-mean bias and data leakage that make results look better than they actually are.
Approach
Built an end-to-end pipeline parsing FastSurfer structural (and perfusion) brain features across a multi-decade aging cohort (n=574, ages 26–91), then iterated through four model versions to systematically diagnose and fix a regression-dilution bias: age-stratified repeated cross-validation, nested hyperparameter tuning with no leakage, inverse-frequency sample weighting to counter age-distribution skew, correlation-based feature selection, and permutation-based (unbiased) feature importance.
Result
Best honest cross-validated model (ElasticNet, combined structural+perfusion features) reached MAE ≈ 5.76 years vs a 7.50-year mean-baseline — a real, permutation-validated signal, reported alongside its actual limitations rather than an inflated accuracy claim.
Featured

More flagship projects

Neural networks — image classification

Image Classification — CNN, LIME & Transfer Learning

PythonTensorFlowResNet50GANLIME
Problem
Understand and visualize how a computer vision model makes decisions, beyond a plain accuracy score.
Approach
Designed and trained CNNs across three datasets (felines, vehicles, landscapes), with cross-validation and GridSearch for hyperparameter tuning. Reduced overfitting via ImageDataGenerator, applied transfer learning with ResNet50, and experimented with GAN/WGAN to generate synthetic images. Added heatmap visualization and LIME explainability, exposed through an interactive site comparing multiple architectures.
Result
An interactive demo for exploring the behavior of several CNN architectures, beyond a static notebook.
Orange sports car misclassified by the model as a tiger Decision-zone overlay showing which regions of the image drove the tiger prediction

A fun real example: the model classified this orange sports car as a tiger — its dominant orange tone and the headlights' eye-like shape were apparently close enough to the tiger class' visual signature.

View demo → GitHub →
Freelance · Computer Vision

Presence Detection System

CNNYOLOPython
Problem
A client needed to automatically identify who was on-site, distinguishing known team members from new entrants — without manual check-in.
Approach
Built a computer vision application using CNN and YOLO for real-time detection and differentiation between known users and new entrants.
Result
A working detection pipeline delivered to the client, developed and deployed as a freelance engagement (COMANCI).
Knowledge Graphs

Semantic Data Integration

PythonRDFSPARQLNoSQL
Problem
Structured data scattered across heterogeneous sources, hard to link and query semantically.
Approach
Built a semantic integration pipeline turning structured data into knowledge graphs: relation extraction, RDF triple generation and SPARQL queries, with a modular architecture for adding new sources.
Result
A reusable, extensible pipeline for turning raw data into queryable graphs.
GitHub →
More projects

Academic & personal projects

Graph Coloring Optimization

Graph coloring algorithms to reduce resource conflicts in distributed systems.

PythonGraph Theory
Details →

Video Game Database

Relational database to track player progress.

SQLPL/SQLPostgreSQL
Details →

Huffman Compression

Compression algorithm to efficiently reduce text data.

PythonAlgorithms

Staffing Agency App

Mobile app to manage assignments and candidate profiles for staffing agencies.

MobileSQLPostgreSQL
Details →

Mem-next — Augmented Memory

Tool to extract key information from text data, feeding an augmented memory system.

PythonNLP
Details →

NLP — Semantic Relations

Extraction of semantic relations from a lexical network.

PythonNLP
Details →

LeCompteEstBon

Reach a target number using six numbers and the four basic operations.

SchemeFunctional
Details →

Tournament Management

Web app to manage tournaments: registration, match generation, scoring, brackets.

HTML/CSS/JSPHPDatabase
Details →