Skip to content

About · Curriculum vitae

Nassim Arifette

Machine-learning research & engineering · Computer vision · 3D · Medical imaging

Download CV (PDF)

01 · Profile

My background includes MVA coursework at ENS Paris-Saclay (17/20) and research internships at CEA, Collège de France, and Université Paris-Saclay. I build reproducible deep-learning systems for 3D vision, medical imaging, reliable ML, and structured data, with careful evaluation of both results and limitations.

02 · Experience

BioMaps, CEA & Université Paris-Saclay • Research Intern (2025)

  • Built a reproducible 3D CT↔UTE-MRI pipeline; trained 3D CycleGAN variants on 311 CT and 292 UTE-MRI training cases.
  • Added a histogram-aware loss → CT→MRI FID 225.96→217.89 (−3.6%), KID 0.1158→0.0969 (−16.3%).
  • Achieved cycle fidelity of ≈23.4 dB PSNR and ≈0.68 SSIM across validation cohorts.

Collège de France, CIRB • Research Intern (Jun 2023 – Aug 2023)

  • Enhanced Phyloformer with triangle-inequality constraints; trained on 1,000 genetic sequences (30 min on V100).
  • Maintained RF-distance while cutting constraint violations from 15% to <1% (trees with 10–100 leaves).

03 · Education

ENS Paris-Saclay — Master 2 coursework (MVA), 2024–2025 • Grade average 17/20 — Deep Learning for Medical Imaging, 3D Vision, Generative Models.

Université Paris-Saclay — Master 1 (AI), 2023–2024 • Grade average 16/20 — NN Verification, NLP, Convex Optimization.

Université Paris-Saclay — BSc (Math & CS, research track), 2021–2023 — Statistical Learning, Algorithms, DB Systems.

04 · Skills

Programming: Python, C++, Julia, OCaml, SQL, Rust, Coq, JavaScript (React, Node.js)

ML/DL: PyTorch, TensorFlow, JAX, MONAI, scikit-learn, Hugging Face

Infra: CUDA, Docker, Git, Linux, SLURM, AWS, MLflow, Hydra/OmegaConf

Languages: French (Native), English (C1)

05 · Selected projects

  • CT → MRI Synthesis: Unpaired 3D image translation with a custom histogram-aware loss.
  • Neural Network Verification via Set Analysis — Explored zones (DBMs) and tropical-geometry prototypes for tighter ReLU-network bounds, evaluated on ACAS Xu.
  • Deep Learning for Voiced/Unvoiced Speech — CNN on MFCCs (85.2% validation accuracy) with 7-fold cross-validation; investigated augmentation for overfitting.
  • YOLOv1 Reimplementation — Trained on Pascal VOC and benchmarked five YOLO variants, with exported weights and demos.

06 · Original document

Preview CV in the browser

The inline preview is hidden on small screens. Open the CV PDF.