Mathias Villerabel
AI Engineer & Data Scientist | Production ML systems, time series forecasting, and billion-record data pipelines | Ex-Swiss Re | Sorbonne CS graduate
Mathias Villerabel — AI Engineer
Production ML for time-series, IoT, and large-scale data pipelines.
- 10⁹+ records/batch — PySpark pipelines at Swiss Re
- 12% forecast lift — demand prediction at Pernod Ricard
- < 2h reconciliation — automated from multi-day manual reviews
- Real-time IoT ML — sensor-to-model pipelines at Equans
I build machine learning systems that ship. Currently at Equans, I own the full pipeline from IoT sensors to production ML — Azure IoT Hub, Kafka, Spark/Polars for distributed feature engineering, model serving and monitoring on Palantir Foundry. Previously built billion-record pipelines and IFRS 17 simulation engines at Swiss Re, and custom PyTorch forecasting frameworks at Pernod Ricard. Before that, 3 years of applied research in Tokyo (SAR/computer vision, published in IEEE Access, 2nd Prize at NEDO Challenge).
Featured Projects
Time Series Forecasting at Scale
Stack: PyTorch, Hydra, MLflow, Optuna, FastAPI, Docker, Streamlit Result: 7.1% MAPE on Walmart M5 (28-day horizon, walk-forward validation) Repo | Demo (~30s cold start)
Production-grade forecasting with 10+ model architectures across retail, rideshare, and inventory domains. Config-driven experiments, tracked runs, Bayesian HPO, containerized serving.
Movie Recommendation System
Stack: PyTorch, Hydra, MLflow, Optuna, Gradio, Streamlit Result: RMSE ~0.323 on MovieLens (100K+ ratings) Repo | Demo (~30s cold start)
End-to-end recommendation system — 4 architectures, full ML lifecycle from training to global inference API on Hugging Face Spaces.
Floorplan Wall Segmentation
Stack: PyTorch, DeepLabV3, OpenCV, Hydra, ONNX Result: 99.4% pixel accuracy (wall-class F1 significantly above baseline despite 96/4 class split) Repo
Computer vision for architectural floorplan segmentation — handling extreme class imbalance with class weighting, morphological postprocessing, and ONNX export. 50ms inference.
Experience
AI Engineer — Equans | Paris | Jul. 2025 – Present
- End-to-end data pipelines: Azure IoT Hub → Kafka → Palantir Foundry streaming, batch on Spark/Polars
- Distributed feature engineering and predictive maintenance models for industrial refrigeration fleets
- Model serving and monitoring — online inference, drift detection, automated retraining
Data Scientist — Swiss Re | Zurich | Sep. 2023 – Jun. 2025
- IFRS 17 simulation engine (cash flows, discounting, CSM) to benchmark EY outputs
- Automated reconciliation (GL ↔ actuarial models), days → < 2 hours per run
- PySpark pipelines in Palantir Foundry scaling to 10⁹+ records/batch
Data Scientist — Pernod Ricard | Paris | Jan. 2023 – Aug. 2023
- Time-series forecasting (ARIMA, gradient boosting, custom PyTorch framework)
- 3-month demand forecasts improved by 12% vs. analyst baseline
Research & Applied Scientist — Synspective | Tokyo | Dec. 2019 – Jan. 2023
- SAR-based construction detection, object detection for maritime trade (2nd Prize NEDO Challenge)
- Cloud-based Earth observation platform on Google Cloud
ML Intern — National Institute of Informatics | Tokyo | 2019
- Action detection models (MobileNetV2 + LSTM) for UAV systems — published in IEEE Access
Education: MSc Computer Science — Sorbonne University, France
Contact
- Email: m.villerabel@gmail.com
- LinkedIn: linkedin.com/in/mathias-villerabel-ba32b0146
- GitHub: github.com/MrEleden