Ibrahim Al Ayoubi
AI Engineer
Specializing in Machine Learning & Computer Vision. Applied research at INRIA & DVRC. Building state-of-the-art models for image segmentation and multimodal AI.
Who Am I?
I build machine learning systems from data preparation through training, evaluation and deployment. My background combines applied computer vision research with the engineering of retrieval systems and data pipelines.
French
C1
English
C1
Arabic
C2
Education
Master’s in Computer Science — Artificial Intelligence & Data Science
University of Montpellier
2024–2026
Double Bachelor’s in Mathematics & Computer Science
University of Montpellier
2021–2024
3
Internships
4
Projects
2
Manuscripts
Technical Arsenal
A broad toolkit for building data-driven models, scalable pipelines, and intelligent systems.
Programming Languages
Python for ML models and data pipelines; SQL and Java for data integration and knowledge graph workflows.
Data Science & AI
Train and evaluate vision models, improve training data with augmentation, and connect language models to retrieved knowledge.
Databases & Big Data
Prepare large datasets, connect relational sources to knowledge graphs, and index embeddings for semantic retrieval.
Tools & Infrastructure
Build REST backends with FastAPI, package services with Docker, and run reproducible experiments with GPU tooling and version control.
Applied in
Soft Skills
Work History
Research and industry internships where I applied state-of-the-art ML & data science.
Machine Learning / Computer Vision Engineer
- Designed and benchmarked DINOv2-based U-Net, TransUNet and DeepLabV3+ for medical image segmentation.
- Built the full pipeline: preprocessing, training, validation, checkpointing, inference and evaluation with Dice, IoU and HD95.
- Evaluated advanced data augmentation strategies to improve robustness and reduce overfitting. The segmentation pipeline achieved an absolute Dice improvement of 3–5 percentage points over the baseline.
Machine Learning & Knowledge Graph Engineer
- Engineered an ETL pipeline ingesting several million weather records for a precipitation-forecasting workflow.
- Designed an ontology and SQL–RDF mappings to unify heterogeneous relational sources into a knowledge graph.
- Implemented existential rules in InteGraal for ontology-based reasoning and forward-chaining inference.
Data Scientist Intern
- Explored predictive modeling for decision support and participated in collaborative data science workflows.
Featured Work
Research and engineering projects showcasing applied ML and computer vision.
Medical image segmentation
CNN, Transformer & Self-supervised Models
Benchmarking CNN, Transformer and self-supervised architectures to improve pixel-level predictions in medical images.
A research library you can query
RAG for Medical Imaging Research
An end-to-end retrieval-augmented generation system that connects a medical imaging paper collection to an LLM through a FastAPI backend.
Climate data, connected
Weather Data & Knowledge Graph Reasoning
A data pipeline and knowledge graph workflow for precipitation forecasting, connecting heterogeneous weather records with logical reasoning.
Mini CLIP
Multimodal Image–Text Encoding
A CLIP-inspired multimodal model that aligns images and text in a shared latent space for cross-modal retrieval.
Publications & Certifications
Research Papers
S2GIM: Superpixel- and Saliency-Guided Image Mixing Data Augmentation for Medical Image Segmentation
N. Kozah1,4, F. Dornaika1,2,*, I. Al Ayoubi3, J. Charafeddine3, F. Davoine5
S2GIM combines superpixel regions and saliency-guided mixing to generate augmented medical images and matching segmentation masks while preserving anatomical contours and salient structures. Experiments on four datasets with convolutional and transformer-based models report improved segmentation accuracy over the compared augmentation methods and lower boundary errors in most cases.
Superpixel structure
Stochastic masks from two source images preserve local region boundaries.
Saliency guidance
Region-level mixing weights average saliency values within homogeneous superpixels.
Unified augmentation
A four-case pixel-wise strategy combines hard and soft mixing in one model-independent process.
Read abstract & affiliations
Deep learning-based medical image segmentation often suffers from limited annotated data, leading to overfitting and reduced generalization performance. Data augmentation is an effective strategy for mitigating this problem by increasing the diversity of training samples. Traditional local image editing methods, such as Mixup and CutMix, generate augmented samples using global interpolation or rectangular region replacement, which may destroy anatomical boundaries and important local structures. More recent superpixel-based approaches preserve contour information but do not fully exploit saliency information within a unified augmentation framework.
To address these limitations, we propose Superpixel- and Saliency-Guided Image Mixing (S2GIM), a novel data augmentation method for medical image segmentation. S2GIM first generates stochastic superpixel masks from two source images and computes saliency-guided region-level mixing coefficients by averaging saliency values within homogeneous superpixel regions. A unified four-case pixel-wise mixing strategy is then applied to produce a single semantically consistent augmented image and its corresponding segmentation mask. Unlike previous methods that separately perform hard and soft mixing, S2GIM simultaneously preserves contour and salient information within one augmentation process.
Extensive experiments conducted on four medical image segmentation datasets, namely ISIC 2017 Task 1, GlaS, MoNuSeg, and Synapse, using both convolutional and transformer-based architectures, demonstrate the effectiveness and generality of the proposed approach. Compared with existing augmentation methods, including Mixup, CutMix, LCAMix, and HSMix, S2GIM improves segmentation accuracy and, in the majority of cases, reduces boundary errors. The proposed method is model-independent, easy to implement, and can be readily integrated into existing medical image segmentation pipelines.
Author affiliations
- 1.University of the Basque Country UPV/EHU, San Sebastian, Spain
- 2.IKERBASQUE, Basque Foundation for Science, Bilbao, Spain
- 3.De Vinci Higher Education, De Vinci Research Center, Paris, France
- 4.Lebanese International University LIU, Beirut, Lebanon
- 5.LIRIS-UMR 5205, CNRS, INSA Lyon, France
Boundary-Aware Superpixel Mixing for Robust Medical Image Segmentation
Research into data augmentation for medical image segmentation, exploring boundary-aware superpixel mixing to improve model robustness.
Certifications
Fundamentals of AI Agents Using RAG and LangChain
IBM
Machine Learning Professional Certificate
IBM
Get In Touch
I am available for full-time AI engineering roles, with a focus on computer vision, machine learning, and applied AI.
Open to Opportunities
Whether you have a research proposal, a job offer, or just want to talk ML & Data Science — my inbox is always open.