Articles
Agentic Engineering — Beyond the Prototype
Coding agents shine in prototypes. In production, they face requirements, project history, and team context they cannot see. Five obstacles explain why the promised 10x gains collide with reality.
- Agentic Engineering
- Software Engineering
Harnessing: 9 practices from the software industry
There is no need to reinvent software engineering to keep coding agents under control: nine established practices can be adapted to working with them.
- Agentic Engineering
- Software Engineering
Software architecture tests: 4 examples to get started
An architecture rule that nobody checks will eventually be forgotten. Here are four ways to make such rules executable as tests run in CI.
- Architecture
- Software Engineering
MLOps: In the run phase, toil is an enemy to face head-on
One ML product can be maintained by hand. By the tenth, the same repeated tasks can bring a team to a standstill: that is toil. Before eliminating it, you have to see and measure it.
With Aurélien Massiot.
- MLOps
- Software Engineering
Architecture Decision Record: let’s start today
Why was this architecture chosen? By the time nobody can answer, the decision should already have been recorded. Markdown, Git, and diagrams as code are enough to get started.
- Architecture
- Collaboration
Summary of Benjamin Bayart’s talk – “Under the hood of sovereign cloud”
“Sovereign” is not a magical property of a cloud platform. This account of Benjamin Bayart’s talk starts with legal, technical, and strategic constraints, then examines the available options without the usual mystique.
- Cloud
- Security
Monitoring & proactive notification of an ML application
An ML application can remain available while its results become steadily less reliable: this guide works back from possible failures to design useful probes and alerts.
- MLOps
- Monitoring
Integrating security early into Machine Learning delivery
Data poisoning, model theft, information leakage: Machine Learning creates its own attack surfaces. Security therefore belongs in the delivery process, not after deployment.
- MLOps
- Security
Cynefin x Machine Learning matrix – Going fast to production to minimize the risk of complex systems
Does deploying early really reduce risk? It depends on the kind of problem being solved. Applied to Machine Learning, Cynefin helps distinguish when to learn from production and when to proceed more cautiously.
- Data Science & AI
- Agility
What if ML monitoring metrics became features?
Monitoring does not have to remain an internal concern. Showing users how fresh the data is or how reliable a prediction may be can turn an operational metric into a product feature.
With Touraya El Hasssani and Antoine Moreau.
- MLOps
- Monitoring
Making the value chain visible in a Machine Learning Delivery project
Mapping the journey from a business idea to a model in production exposes what delivery metrics often hide: queues, rework, and manual tasks.
With Guillaume Pivette.
- MLOps
- Agility
Business and Data Scientists must collaborate to design reliable algorithms
A reliable algorithm is not built in a private conversation between a Data Scientist and their data. The business must also define the objective, the trade-offs, and which errors are acceptable.
- Data Science & AI
- Collaboration
Test data management in Machine Learning delivery
Reliable tests require test data that teams can actually control. This article compares several ways to obtain it on demand without sacrificing relevance or confidentiality.
- MLOps
- Software Engineering
Visual management in a Machine Learning Delivery project
How do you make work built around experiments, training runs, and annotation visible? By adapting visual management to Machine Learning delivery instead of imposing a standard Kanban board.
With Maria Mokbel.
- MLOps
- Agility
Ten convictions for more successful Data Science projects in 2021
In 2021, we set out ten convictions drawn from our Data Science projects. The common thread remains straightforward: deliver early, put models in front of users, and work closely with the business.
With Eric Biernat.
- Data Science & AI
- Collaboration
The challenge of obtaining ground truth
In production, ground truth may remain unknown, arrive too late, or depend on user feedback. Without it, there is no way to know whether the model is still achieving its goal.
- MLOps
- Monitoring
An alternative to distribution monitoring
Monitoring every distribution often amounts to collecting charts. A better approach starts from uncertainty in the pipeline and places a few probes where failures can actually be observed.
With Mehdi Houacine.
- MLOps
- Monitoring
Interpretability of Data Science Systems
“This model must be interpretable.” Fine, but to whom, for which decision, and at what cost? Asking those questions avoids the search for a universal explanation that does not exist.
- Data Science & AI
Benchmark of dataPreparation versus standard R
For small datasets, standard R does the job. As the data grows, this benchmark measures how much time and memory dataPreparation saves when centering and scaling variables.
- Data Engineering
- Software Engineering
On data and quality
Data quality problems often begin long before anyone uses the data; tracing them back to their source reveals five concrete ways to reduce them.
- Data Engineering
- Data Science & AI
AI, the end of employment?
Will AI end employment, or change what we mean by work? This essay looks back at earlier technological revolutions to place today’s fears in a longer history.
- Data Science & AI
- Society