Portrait of Abdennour Kerboua

Abdennour Kerboua

Paris, France

Research Engineer @ Idemia

[ vision researcher · humanist · gamer / reader · optimist ]

contact@abdennour.tech|abdennour.kerboua.pro@gmail.com|résumé ↗

Experience

Apr 2026 - Present

Courbevoie, France

Computer Vision Research Engineer @ Idemia

person and luggage re-identification, multi-view and 3D modules

Computer VisionRe-IDDeep Learning

Aug - Sep 2024

Paris, France

Frequency Control Technician @ ANFR

spectrum monitoring for the Paris 2024 Olympic and Paralympic Games

RFMonitoring

Projects

Education

2025 - 2026

Gif-sur-Yvette, France

Master in Mathematics, Vision, Learning

ENS Paris-Saclay

2023 - 2026

Paris, France

Engineering Degree, GPA 4/4

Télécom Paris, Image Processing, Signal Processing, Artificial Intelligence

Notes

Playlist

Evidentia

Local agentic RAG for scientific papers: fact planning, Qdrant retrieval and reranking, citation-controlled answers, and abstention when evidence is insufficient.

Final v6, 30 cases: 29/30 end-to-end passes, 100% expected-document retrieval recall, 100% citation precision and recall, 86.1% factual coverage.

Lire le détail ↗ GitHub ↗

ReferAI

Computer-vision pipeline that turns football video into a structured game state for future refereeing assistance.

Current

Next

GitHub ↗

Drawing Clouds

Reinforcement-learning sandbox where a policy learns to shape a 2D cloud into a target drawing through local weather actions.

Drawing Clouds V3: target drawings, initial weather states, and generated clouds across fifteen shape families
V3 evaluation board, best local checkpoint at 650k steps.

Approach

Result

Best V3 checkpoint: 0.731 mean score, 0.727 precision, and 0.469 boundary alignment. Filled shapes are strongest; hollow contours remain the main challenge.

GitHub ↗

Molecular Graph Captioning

Graph convolutional networks, SciBERT and contrastive learning for molecular graph and scientific-text alignment.

Lire le détail ↗ GitHub ↗

Self-supervised Medical Imaging

SimCLR and Barlow Twins on MedMNIST, then downstream evaluation with limited labelled data.

Lire le détail ↗ GitHub ↗

IBM x Telecom Fraud Detection

Feature engineering and Random Forest classification across 300k transactions. F1: 69%, third of 25 teams.

Lire le détail ↗ GitHub ↗

Sign Language Recognition

MediaPipe hand landmarks classified by a Random Forest from live camera input.

Lire le détail ↗ GitHub ↗

Exemplar Patch Inpainting

Criminisi inpainting: priority-based boundary filling and exemplar patch matching for object removal.

Lire le détail ↗ GitHub ↗

QR Navigation Car

Raspberry Pi vehicle using camera-based QR detection as visual navigation waypoints.

Lire le détail ↗ GitHub ↗