cv

CIFRE PhD Candidate — Tabular Foundation Models & Generative AI

Orange Innovation / IRISA - INRIA

PhD program in partnership with Orange Innovation Labs and the IRISA Research Institute.

• Designed and trained from scratch a foundation model for structured/tabular data (<100M parameters) — full pipeline: architecture design, pre-training corpus curation, training at scale, evaluation on unseen schemas
• Designed an original Datum-wise Transformer architecture from scratch for heterogeneous tabular data — +7 AUC points and +7 accuracy points over the only published baseline, with only 3M parameters (vs. 110M+); extended via domain adaptation (gradient reversal) for cross-schema generalisation: additional +7 accuracy points
• Fine-tuning and adaptation of LLMs on structured data — tabular serialisation, heterogeneous feature encoding, transfer to unseen schemas
• Rigorous benchmark of 5 generative model families (VAE, GAN, Diffusion, LLMs, classical) on 16 datasets with 300 HPO trials per fold — end-to-end pipeline: data curation, model design, training, evaluation and ablations
• Large-scale distributed training on HPC clusters (Jean-Zay, IGRIDA, Marcel/Orange) — Slurm job management, Tesla V100 32GB and RTX 4090 GPUs
• First author on 4 international publications: AAAI 2026 (A* conference, top 3 worldwide in AI), Neurocomputing Q1, IDA 2025, COLING 2025; invited talk at CFE-CMStatistics 2025 (Birkbeck University of London, ~1 800 attendees)
• Co-supervised an M2 intern (6 months) on foundation model extension — topic definition, weekly technical mentoring, and result evaluation

Apprentice — Software Developer & Data Scientist

Kurmi Software

Work-study program as part of my master's degree in big data.

• Developed Kurmi connectors for integration with Microsoft 365 (Java, JavaScript, PowerShell)
• Behavioral analysis of Kurmi logs across 21 dimensions (67 operation types, 19 entity types) for user segmentation — Python/KNIME pipeline, Power BI visualization
• Designed and maintained software tests to ensure the quality of the Kurmi suite

Research Intern — DRUID Team

INRIA - IRISA

Trainee to validate my Bachelor's degree.

• Analyzed the evolution of scientific fields through NLP co-occurrence graphs on 22 240 arXiv papers (cs.AI, 1993–2019) — graphs up to 9 500 nodes and 2.8M edges
• Distributed computing on the IGRIDA grid (35 nodes, 696 cores) for graph generation and spectral analysis

Artificial intelligence intern, then technical management assistant

Machine Intelligence For You (MIFY)

Post-graduate work experience in machine learning.

• Collected data using a graphics tablet to support a machine learning algorithm for forged signature detection
• Jointly delivered a Microsoft Excel training program within SUNU Assurances
• Facilitated brainstorming sessions and managed team coordination
• Improved the company's website SEO and page loading times

PhD

University of Rennes, Rennes, France

Computer Science

• Synthetic Tabular Data: Generation and Detection

Bachelor's & Master's Degree (MIAGE)

University of Rennes, Rennes, France

Computer Science

• Training professionals to lead large-scale projects at the intersection of business management and IT, leveraging databases and big data technologies.

Robust Detection of Synthetic Tabular Data Under Schema Variability

AAAI 2026 (A* conference, top 3 in AI), Singapore

Proposed a novel detection method for AI-generated tabular data, employing a Datum-wise Transformer architecture that achieves +7 AUC and +7 accuracy points over the baseline with only 3M parameters.

Invited talk: "Synthetic tabular data detection in the wild"

CFE-CMStatistics 2025, Birkbeck University of London, London, Dec. 2025 — International hybrid conference ~1 800 participants

Invited talk in the session "Synthetic data: Generation and validation methods" at an international hybrid conference in econometrics and computational statistics.

Synthetic Tabular Data Detection In the Wild

International Symposium on Intelligent Data Analysis (IDA), Konstanz, Germany, May 2025

Employed state-of-the-art tabular data generation models to evaluate encoding strategies and detection methods for AI-generated content.

Cross-table Synthetic Tabular Data Detection

Workshop GenAIDetect, COLING 2025, Abu Dhabi, Jan. 2025

Developed baseline models for synthetic tabular data detection with a cross-table shift (deploying on unseen tables).

Tabular Data Generation Models: An In-Depth Survey and Performance Benchmarks with Extensive Tuning

Neurocomputing (Q1 Journal, Elsevier), 2025

Conducted an extensive benchmark of generative models for tabular data, optimizing hyperparameters, feature encodings, and architectures across 16 datasets with 300 HPO trials per fold.