Felipe Vicentin

Image Engineering Research Intern · DxO Labs

Boulogne-Billancourt, France

• Researched learning-based multi-frame fusion for joint denoising and demosaicking of image bursts within DxO's research lab, developing code with PyTorch (CUDA), NumPy, OpenCV, and matplotlib.
• Built a synthetic burst-generation pipeline in RGB and RAW, turning a public image dataset into multi-frame bursts with ground truth through reverse-ISP unprocessing, Bayer mosaicking, simulated handheld motion and a shot/read sensor noise model.
• Wrote a batched coarse-to-fine GPU implementation of ECC fast enough to run inside the training loop, and derived a reference-free alignment-quality metric from a noise-normalised chi-square statistic.
• Built the synthetic burst-generation pipeline behind every training (multi-frame bursts with ground truth, from unprocessing and Bayer mosaicking to simulated handheld motion and sensor noise), and wrote a batched GPU implementation of the ECC alignment algorithm fast enough to run inside the training loop.
• Trained U-Net burst models with a controlled ablation (single-frame, unaligned, oracle and estimated alignment): aligned multi-frame fusion gained 3-5 dB PSNR over an identical single-frame baseline while the unaligned model matched it, isolating registration quality rather than model capacity as the source of the gain.
• Full-time position (35 h/week) over 6 months, equivalent to the final master's research placement in the French system.

Undergraduate Research Fellow (FAPESP-funded) · Unicamp, IMECC

Campinas, Brazil

• Researched partial orders and lattices over finite chains, advised by Prof. Peter Sussner (Department of Applied Mathematics).
• Developed rigorous proofs on counting and constructing admissible orders of intervals over finite chains, published as "On counting and constructing all admissible orders of the non-empty intervals in any finite chain", Fuzzy Sets and Systems, 2025 (DOI: 10.1016/j.fss.2025.109372).

On counting and constructing all admissible orders of the non-empty intervals in any finite chain

Data Engineering Intern · Dextra Systems

Campinas, Brazil

• One-year internship at a large IT consultancy (~1,000 employees), as the only intern in an agile squad delivering to enterprise clients across sectors ranging from food to banking.
• Built PySpark and Delta Lake pipelines on a layered data lake (ingestion, anonymisation for LGPD compliance, standardisation, KPIs) over conversation data, running on GCP with Airflow, Cloud Storage, BigQuery and Dataproc across separate production and staging environments.
• Delivered production pipelines for 4 enterprise clients and 15+ KPIs, including an ML layer running LDA topic modelling over cleaned text.

Master of Science (M2), MVA (Mathematics, Vision, Learning)

Université Paris-Saclay, ENS Paris-Saclay · Paris, France

• Selective research master's in mathematics, vision and learning, taken as a double degree between Unicamp and IP-Paris.
• Representation Learning, Convex Optimization, Optimal Transport, Probabilistic Graphical Models, Geometric Deep Learning, Kernel Methods.
• Generative Models for Images, Large-Scale Model Training and Deployment, Advanced Deep Learning, Time Series, Reinforcement Learning.

Subjects I have written about

• optimal transport
• convex optimization
• functional analysis

Master of Engineering (Diplôme d'Ingénieur)

Institut Polytechnique de Paris, Télécom Paris · Palaiseau, France · Grade 16.3 / 20

• Alumni excellence scholarship.
• Linear Models, Machine Learning, Variational and Bayesian Methods and Discrete Optimization, Image Learning and Object Recognition, Statistics, 3D Vision and Video.

BSc in Computer Science

University of Campinas (Unicamp) · Campinas, Brazil · Grade 9.34 / 10

• Calculus, Linear Algebra, Discrete Mathematics, Databases, Algorithm Analysis, Data Structures, Operating Systems, Introduction to Cryptography.

vicentin

• Python library with a dual-backend architecture that transparently switches between NumPy and PyTorch based on input data, enabling automatic differentiation and GPU use with no API change.
• From-scratch implementations in optimization (Newton, barrier, proximal/ISTA), machine learning (SVM, PCA, kernel methods) and deep learning (VAE, Wasserstein GANs), plus paper reimplementations (Grad-CAM, SimCLR).

Multi-categorical LAFTR

• Implementation of the LAFTR model, extending it to multi-categorical settings. Supervised by Prof. Loïc Le Folgoc.
• Learns representations that stay predictive while making a protected attribute hard to recover — the core problem behind fairness auditing in credit and risk decision systems.

Image restoration

• Implementation of variational techniques (Tychonov and Total Variation) to invert transformation kernels on noisy images, using NumPy and PyTorch. Supervised by Prof. Arthur Leclaire.