I build AI that
catches fraud &
reads sentiment.
AI engineer at Groupe Covéa, Paris. I lead fraud-detection models end-to-end — recovering millions of euros in claims — and build sentiment-analysis systems with fine-tuned LLMs that sharpen customer intelligence. Fine-tuning, LoRA adapters, prompt engineering, production monitoring: I own the full stack.
Expertise
Fraud Detection
Production fraud-scoring models for insurance claims at Groupe Covéa — several million euros recovered at MAAF, 75% alert precision, 3× faster alert handling. I lead the full lifecycle: framing with fraud experts, engineering, deployment, monitoring.
XGBoost · scikit-learn · MLflow · GrafanaSentiment Analysis
Domain-specific sentiment systems: fine-tuned CamemBERT & Mistral with LoRA / DoRA adapters, generative-AI pipelines for customer intelligence, and a working paper benchmarking FinBERT & CryptoBERT for crypto-market forecasting.
CamemBERT · LoRA / DoRA · FinBERTLLM Engineering
End-to-end LLM pipelines in production — fine-tuning, benchmark frameworks across 40 labels, vLLM serving, and annotation workflows accelerated up to 100×.
Mistral · vLLM · LangChain · MLflowAsk
Stack
Experience
Data Scientist—Groupe Covéa
- —Leads the full lifecycle of fraud detection models (framing → engineering → deployment → monitoring), recovering several million euros at MAAF through highly predictive scoring.
- —Develops fraud models for civil liability at GMF and home insurance at MAAF, strengthening claim controls and reducing financial leakage.
- —Achieves up to 3× faster alert processing through close collaboration with domain fraud experts.
- —Builds client sentiment-analysis systems using generative AI and advanced prompt engineering to enhance customer intelligence and risk signals.
Machine Learning Engineer—Groupe Covéa
- —Fine-tuned Mistral, CamemBERT, and GPT with custom LoRA / DoRA adapters for domain-specific classification and text generation.
- —Industrialized text classification — cut annotation time from ~90 s → instant, accelerating the overall process by up to 100×.
- —Achieved ~80% average performance across 40 labels via benchmark frameworks covering accuracy, latency, and robustness.
- —Built end-to-end ML pipelines with MLflow for experiment tracking, reproducibility, and deployment.
Data Scientist—Octopize
- —Implemented a comprehensive evaluation pipeline for avatarized (anonymized) data performance across multiple ML models.
- —Automated testing for scalable evaluation; developed benchmarks comparing avatarized datasets against original data.
- —Contributed to a research article on innovative avatarization approaches.
Team Leader—Centrale Nantes
- —Led a multidisciplinary team end-to-end to deliver a supercapacitor-powered bike — managing technical design, budgeting, and client requirements.
- —Acted as the primary client liaison, translating weekly feedback into engineering solutions aligned with performance and cost goals.
Web Developer—InMobiles Holding S.A.L.
- —Built full-stack web applications with Django and maintained REST APIs via Django REST Framework.
- —Enhanced system security with OTP and reCaptcha authentication.
Internship Trainee—OGERO
- —Studied XDSL, FTTX, PON, and ODN infrastructure at Lebanon's national telecommunications operator.
- —Located cable faults and tested FTTH networks using OTDR.
Teacher—Self-employed
- —Taught math and physics to many students over ≈ 6 years.
Work
X-code EVAL
Full MLOps pipeline classifying coding exercises into the correct label — end-to-end from data ingestion to deployment.
Mandelbrot & Julia Sets
Generation and visualization of Mandelbrot and Julia fractal sets.
Education
Engineering Degree in Artificial Intelligence—Centrale Nantes
Diplôme d'Ingénieur · Grande École · double-degree track
Engineering Degree in Telecommunications—Lebanese University ULFG2
Electrical, Electronics & Communications Engineering · double-degree track
Contact
Let's
Build.
Open to ML / LLM engineering roles and freelance projects. Let's make something that ships.