Train your teams in modern Python data science
From prototype to high-performance production
Intensive in-company training in Python, data science and high-performance computing. For data scientists, engineers and researchers in companies.
In English or French — on site or remote
Three data science courses designed for production
Each course lasts 3 days (21 hours), on site or remote, for 10 participants at most.
Python data science — level 1
Data analysis and visualization
- NumPy
- Pandas
- Seaborn
The foundations of the Python data science ecosystem. Vectorization, broadcasting, DataFrame manipulation, time series, the Parquet format, statistical visualization with Seaborn.
See the program →Python data science — level 2
Advanced manipulation and high performance
- Polars
- Numba
- Multiprocessing
Going beyond Pandas when data exceeds RAM or when performance matters. Polars and its streaming engine, JIT acceleration with Numba, multi-core parallelism with multiprocessing.
See the program →Python machine learning
From concept to production
- scikit-learn
Hands-on mastery of supervised and unsupervised machine learning with scikit-learn. Pipelines, regularization, ensemble methods, model selection. Emphasis on methodological rigor: avoiding overfitting, cross-validation, selection without data leakage.
See the program →Do your teams have the required Python level?
The data science courses assume a good command of Python: functions, iteration, modules, namespaces. If this is not the case for all your staff, I can add the Python — level 1 module before the data science days. For teams that develop complex code, the Python — advanced module (object-oriented programming, decorators, metaclasses) is also available.
These complements are offered case by case — feel free to mention them in your quote request.
Why trust me with training your teams?
I teach what I practice as a researcher
As a senior research scientist at Inria, I have been using Python for more than 20 years in my own research, on data at very large scale: a complete analysis of the Twitter social graph (505 million accounts, 23 billion links), processing of the entire Bitcoin transaction graph (16.5 billion links, 1.5 billion addresses), massive measurements of exposure to radio waves. This work was covered by the New York Times, the Washington Post, Fortune and Le Monde. When I teach Polars, the streaming engine or the pitfalls of parallelism, these are the tools I use every day.
A teaching method proven on 150,000 learners
Co-creator in 2014, with Thierry Parmentelat, of the Python MOOC on the FUN platform, followed by more than 150,000 learners and used by Sorbonne Université and CentraleSupélec. The material and the learning path have been tested on an extremely varied audience, from complete beginners to experienced developers who want to consolidate their foundations. This diversity taught me to quickly identify a team's real level and to adapt the pace accordingly.
A focus on performance, always state of the art
What sets my catalog apart: going beyond introductory data science. Polars vs Pandas on data that exceeds RAM, CPU-bound parallelism, JIT acceleration with Numba, fine-grained memory management. These are the topics that separate a team that prototypes from a team that ships to production. The material evolves continuously to follow the ecosystem — new Polars versions, changes in the streaming engine, emerging best practices — because this is also my research ground every day.
How I teach
My experience as a researcher, university teacher and corporate trainer has taught me one thing: the quality of a training course is not measured by the number of slides covered, but by the participants' ability to apply what they learned to their own code.
I build my courses around hands-on exercises in Jupyter notebooks, with use cases close to the real problems participants face. I make sure each concept is understood before moving on to the next, even if it means adapting the pace to the group. I remain available for occasional questions in the weeks following the training: learning does not stop on the last day.
Agentic AI
Why train your teams in the age of agentic AI?
Claude Code, Codex, Cursor, Gemini CLI, Aider: coding agents are transforming the engineering profession. A developer can now produce hundreds of lines of NumPy or Polars without mastering the fundamentals. This is precisely what makes training strategic today.
Models are structurally behind the libraries
Training a frontier model takes months, and data science libraries evolve constantly — new Polars APIs, changes in the streaming engine, changes in scikit-learn's behavior, new NumPy releases. Agents often generate code that looks plausible but relies on obsolete APIs, deprecated patterns, or ignores recent optimizations. Without a team able to spot these gaps, the code drifts silently.
Agentic AI amplifies gaps as much as skills
For an engineer who masters the subject, it is a real productivity booster: they know how to steer the agent, check its output and correct its drift. For a team that does not master the fundamentals, it is the opposite: a sink of unmaintainable code, where subtle, hard-to-fix bugs pile up. A recent CodeRabbit study, reported by Stack Overflow, shows that AI-generated code contains 1.7 times more bugs than human-written code, including 75% more logic and correctness issues — exactly the kind of errors that slip through code review and end up as production incidents.
My courses include the use of AI agents
I use these tools myself every day in my research work — Claude Code, GitHub Copilot and others — and I have a cross-cutting view of them. Beyond the Python and data science fundamentals, I teach how to get the most out of coding assistants in a professional setting: structuring requests to get quality code, reviewing and auditing the output efficiently, spotting suspicious patterns, and combining the agent with good testing and profiling practices. The goal: make AI a real productivity lever, while guaranteeing the maintainability and reliability of the code you ship.
Training your teams today gives them the critical skills that turn an AI agent into a real productivity lever, and a lasting competitive edge over teams that endure AI instead of steering it.
Beyond training: an independent look at your data science
Are your data pipelines slow or too expensive? Do your data scientists code with AI agents without anyone measuring what it changes? Does an important decision rest on an analysis that nobody has reviewed? I work as a consultant on these three topics: a one- to two-day diagnostic, an in-depth engagement if it is justified, or ongoing advisory of half a day to a day per month.
Discover consulting →Participant satisfaction
Out of 289 evaluations collected over 39 training sessions since 2022:
Testimonials
“Mr. Legout has a perfect command of the subject. He humbly answers every question in the clearest and most honest way. I learned Pandas and Seaborn very well, and also many very useful things about NumPy: the most efficient methods among several, the way NumPy ‘thinks’. I would gladly attend another course if he gives one.”
“Arnaud Legout is a very good trainer, he explains things very well, knows his subject very well and answers all our questions. The CoCalc tool is very practical: you write code while the trainer can see what you are doing, so he can help when you are stuck. I learned many things that will be very useful in my work.”
“Very well documented. Perfectly meets my needs and goes further. Pedagogical course. The rhythm is fine. The answers to questions are great.”
They trusted me with training their teams
Tech and industry
Orange · Cisco · Intel Mobile Communications · Gemalto · EPCOS · Transvalor · Supersonic Imagine · CNEDI-CERTIAM · Eau d'Azur
Public research and higher education
CNRS · INRAE · INSERM · Observatoire de la Côte d'Azur · Université Côte d'Azur
Specialized training
École des Marins-Pompiers de Marseille
Frequently asked questions
Format and organization
Is the remote format suitable for this kind of training?
Yes, and it is a fair question for such technical courses. Remote training works well here for several reasons: the hands-on exercises are done on CoCalc, a collaborative cloud environment that lets me see each participant's code in real time and step in directly when someone is stuck — exactly as in person. Sessions are limited to 10 participants to ensure individual follow-up. The pace is adapted, with frequent time for questions. In our post-training surveys, 90% of participants rate the remote format as equivalent or better than in-person.
How long are the courses and how are they organized?
Each course lasts 3 days (21 hours), in continuous sessions. The in-company format lets you adapt the daily schedule to your constraints (3 consecutive or spaced-out days). The group is limited to 10 participants.
Can the training take place on our premises?
Yes. The courses are offered on site or remotely. On site, each participant needs a computer with Internet access and a large screen that comfortably fits two windows side by side, or a dual screen. The technical environment (CoCalc) requires no installation.
Content and level
Do participants need to know Python already?
It depends on the course. The Python data science course (NumPy, Pandas, Seaborn) assumes some prior Python experience but no data science. The high-performance data science course (Polars, Numba, parallelism) is for teams already comfortable with the data science ecosystem. If your team needs a Python refresher before starting, I also offer a dedicated Python level 1 course. A preliminary discussion lets us check that the program matches the team's level.
Can the courses be adapted to our team's context?
Yes. The in-company format makes it possible to adjust the content to your use cases, your data and your technical constraints. A scoping discussion before the training identifies the priorities and steers the exercises accordingly.
What is the difference between the three courses in the catalog?
The three courses cover distinct levels and goals: Python data science lays the foundations (NumPy, Pandas, visualization); high-performance data science in Python moves on to performance at scale (Polars, Numba, parallelism, memory management); the third course focuses on machine learning with scikit-learn, from prototype to production. They can be taken independently, depending on the team's level.
Organization and pricing
Are the courses available in English?
Yes. All courses are available in English or French, depending on your team.
How do I get a quote?
Just fill in the contact form. I take into account the number of participants, the format you want (on site or remote), the language and any need to adapt the program. A quote is sent promptly after a scoping discussion.
Can we get references from past training sessions?
These courses have been delivered to teams in many organizations, in companies as well as in public research: Orange, Cisco, Intel Mobile Communications, Gemalto, EPCOS, Transvalor, Supersonic Imagine, Eau d'Azur on the industry side; CNRS, INRAE, INSERM, Observatoire de la Côte d'Azur, Université Côte d'Azur on the research and higher education side. Detailed testimonials are available on this page, and I can provide further references on request.
Do you offer follow-up after the training?
The courses are designed to be immediately applicable. The complete material (notebooks, exercises, solutions) is given to each participant. For specific follow-up needs or longer support, feel free to mention it in your quote request.
Let's discuss your training needs
Each course can be adapted to your teams' context: starting level, business use cases, choice of modules. Describe your needs and I will suggest a suitable format and quote.