Data Scientist · ML Engineer · Paris · GCP Certified
Adrien
Morel
Applied mathematics engineer moving from forecasting and anomaly-detection models into agentic AI and LLM tooling — and the pipelines that keep both running.
From applied mathematics to production machine learning — customer-experience NLP at Transdev, track-geometry anomaly detection at SNCF Réseau, then end-to-end forecasting systems of my own on GCP, and autonomous agents built on LangGraph and MCP. Reserve officer data scientist with the French Navy.
Selected Work
Freelance
01/2025 — present
Data Scientist / ML Engineer · Independent practice
Two strands: a client transformation mission (Espace Trinité, below) and end-to-end ML systems of my own — feature pipelines and feature tables on BigQuery, training and batch-inference pipelines, model registry, CI/CD, drift and data-quality monitoring — alongside agentic AI on LangGraph and MCP, where the hard part is traceability: every claim an agent produces stays tied to its source.
→ Open source contributor (Skore / Probabl) ↗ · Kaggle top 2% worldwide
Espace Trinité
06/2025 — 05/2026
Digital Transformation Consultant · Freelance mission
End-to-end digital transformation of a B2B venue: network and hardware overhaul, software stack migrations, automated operational workflows, and training non-technical teams on digital tools and best practices.
French Navy / iReMMO
09/2023 — present
Reserve Officer (Data Scientist) · Executive Training
Data scientist in the French Navy digital reserve. Executive staff programme (PMS État-Major) and advanced geopolitics training on the Middle East (iReMMO Institute) through 2023–2024, then an operational deployment aboard the aircraft carrier Charles de Gaulle (mission La Fayette 26, 05–09/2026), working alongside operators in a constrained environment at an operational tempo.
→ Ranked 2nd / 180, French Navy staff programme (PMS État-Major)
→ Active Secret clearance
SNCF Réseau
10/2022 — 05/2023
Data Scientist · Railway Infrastructure
Automated anomaly detection on railway track-geometry data, with no absolute ground truth to train against — which ruled out standard supervised learning and led to a signal-processing approach: cross-correlation registration of successive measurement runs, then drift indicators (growth rate, defect clustering) and interpretable visualizations built for maintenance technicians.
→ Cut manual analysis time by 50% for a team of 12 technicians, and reduced critical detection errors
Transdev
06/2021 — 09/2022
Data Scientist / Data Analyst · Urban Mobility · Apprenticeship
Designed single-handedly the group's operational performance reporting tool, centralising every subsidiary. Built an NLP pipeline monitoring customer experience on Twitter — extraction, cleaning, sentiment classification — to surface weak signals during field incidents. Power BI dashboards tracking fuel consumption and driver performance across 3 business units; historical data migrated to Snowflake.
→ Excel/VBA benchmarking tool deployed to 150 users
Open Source & Projects
I run one project in production at a time, and retire it once it has proved what it needed to prove — cloud spend is a real constraint, not an afterthought. Each entry below states the window it actually ran. vigie-01 is next in line for deployment.
Vélib' Forecast
Forecasting · In production 12/2025 — 02/2026
Bike-availability forecasting across ~1,500 Vélib' stations, which ran unattended for three months on a serverless GCP architecture. Vélib' and weather feeds ingested every 5–10 min, temporal/spatial/weather feature engineering, one XGBoost model per horizon, daily retraining. Four monitoring layers: data health, drift (PSI and Kolmogorov–Smirnov), model performance against a persistence baseline, and network dynamics.
→ ~15% lower MAE than a persistence baseline at the 60-minute horizon
vigie-01
Agentic AI · Defense
Autonomous OSINT agent for defense/geopolitical monitoring. LangGraph multi-agent pipeline — collection, classification, verification with bounded tool-calling — FastAPI + React, per-claim source traceability.
elec-forecast
Forecasting
24-hour electricity demand forecasting across 12 French regions on GCP. LightGBM on a BigQuery feature store, MLflow on Cloud Run, CI/CD, drift and data-quality monitoring. Ran a full production cycle, then decommissioned 08/2026 — the case study covers what it cost to operate and why it was retired.
skore
Open Source Contribution
Contributor to skore — the scikit-learn collaborative ML project management library.
mcp-ollama-agent
Agentic AI
Fully local agentic AI system: a custom MCP server, ReAct orchestration, RAG retrieval and on-device Ollama inference. Because the agent holds real action capabilities — code execution, file access — permission scoping and guardrails were designed in from the start rather than bolted on.
Stack
Modelling
scikit-learn, LightGBM, XGBoost, Optuna, SHAP
Statistics
Statistical modelling, spectral analysis, time-series forecasting, anomaly detection
Data
Python (pandas, NumPy, SciPy), SQL, BigQuery, PostgreSQL, Snowflake
LLM & Agents
LangChain, LangGraph, MCP, ChromaDB, Ollama, OpenAI & Anthropic APIs
MLOps
MLflow, Vertex AI Pipelines, Docker, Git, GitHub Actions, drift & data-quality monitoring
Serving & Reporting
FastAPI, Pydantic, Cloud Run, React, Power BI
Certifications
Education
CY Tech
09/2019 — 06/2022
Cergy, France
Cergy, France
Engineering degree in Applied Mathematics. Dual degree with the University of A Coruña (Spain) in High-Performance Computing.
Languages — French (native) · English (fluent, TOEIC 950) · Italian (conversational)
Available from October 2026.