writing
Technical articles and reading notes.
Each article covers a single subject. Most start from a case met during an engagement; the others from a reading or a question of method. All are dated to when the work took place.
01contents
subject
year
period
2020202120222023202420252026
2026The grain of randomness: what decides a repetitionA generator of repeated fields is usually controlled by a single number, the quantity of randomness. Yet that is the less important parameter. What the image looks like depends mainly on the level at which the randomness applies, and that level appears nowhere in the code.2026Gale-Shapley at the milonga: stable does not mean fairThe algorithm recognised by the 2012 Nobel prize guarantees that no pair of dancers would prefer to form outside the matching produced. It guarantees nobody a dance, and the gap between those two properties is larger than it seems.2026Mathematics has no compilerFour mathematicians have started a blog about what AI changes in their work. The discipline believes it is protected by proof and peer review. Yet it works on the same object as software, built step by step, but without the tools that, in software, make rejecting an error effortless. That ease of rejection, more than trust, explains how much I now get done with an assistant.2026AI Act, Cyber Resilience Act, Résilience bill: the same missing inventoryThree European texts fall on the same calendar, with opposite effects on demand: one has just been postponed by eighteen months, another applies in six weeks, the third holds budgets back until it is passed. What they require, however, is very similar.2026Astro: rules before pages, then publishingThe full process, private repository included: design system first, then implementation, then automatic deployment on every push. With the five traps that cost me time because none of them is in the documentation.2026Seeing a drawing inside outOld drawings separated from their paper, then inverted. What the paper/ink switch brings out in a sixteenth-century sheet, and what it distorts, since an analysis tool also changes what it shows.2026Astro: from a CV to a static siteThe full path, for an independent worker or an association: turning a profile and an idea for a logo into a site one person can maintain. The code matters more than the hosting.2026Time saved, strain, and who pockets the difference“It saves me time” is the most common sentence about AI at work, and the least precise, because it combines three claims. The first has a mathematical ceiling, the second can reverse, and the third does not depend on you.2026Agent Skills, out of the clipboardThe instructions given to assistants end up where nothing is reviewed: in conversations and personal files. Getting them out only requires treating them like the rest of the code.2026Neither all salary nor all dividendsThe question a company director settles once a year, usually by guesswork. The answer lies at neither extreme: the curve has a maximum, which moves with the result.2026You are not taxed on what you withdrawThe question fits in one sentence: how much can I live on without drawing down my capital? In France, the answer runs into a misunderstanding that is costly either way: people expect to pay the headline rate, which applies to only part of the withdrawal.2026Two disciplines of lookingChess and drawing teach the same thing by two paths: recognising a configuration before calculating, and calculating when recognition is not enough.2026Mathematics seen from the insideA research diary, a thesis on intuition, the history of a single number: three books that show mathematicians at work, not only their results. And five others, outside computing.2026TMDL in Fabric: two routes for the same fileNow that a semantic model is written as text, it can be reviewed and versioned like code. There remain two ways of deploying it, with different costs, suited to different teams.2026Through the tradesman’s entranceA bitter memoir, a television series and a comic: three ways of explaining science to people who will never read a proof, and what each one leaves out.2026dbt exposures generated from TMDLThe dependency graph stops at the last table, while the question asked in meetings is about what comes after. A single repository and a text format are enough to extend it, provided the extension is generated and never written by hand.2025Choosing engagementsTwo long engagements, and time between them to examine the practice itself: what rate, what to refuse, which tools to keep for ten years.2025dbt-fabric called from a Fabric notebookAn integrated platform does not always provide somewhere to run a dbt project. dbt’s Python entry point provides one, provided you accept a few peculiarities.2025What is decided before the lineThree books about what is decided before the line: the viewpoint, the volumes that make up the form, and the distribution of dark masses.2024On what horizon to judge a systemA year spent moving from one client to another, seeing up close how an AI budget line is approved or rejected in committee, and over what horizon a system should be judged.2024PyMuPDF and pgvector: what the chunking decidesIn a retrieval-augmented pipeline, attention goes to the model. Yet quality depends on two earlier steps: how a PDF is split, and what the extraction has already lost.2024Polars and Delta from a Durable FunctionA transactional table format and a function you do not control fit together badly: the first assumes a stable writing process, the second can be replayed at any moment. What that requires, and where the limit lies.2023Deciding whether to buildThe year generative AI moves from technology watch into meetings, where the question is often badly framed. And the year my role in those meetings changed.2023dbt under Airflow: one task per model, not one per projectA dbt project launched from an orchestrator usually sits in a single task. The graph displayed then does not match the project, and a retry after an error reruns everything. What changes when the graph is handed to the orchestrator, and what it costs.2023SageMaker and EMR from Airflow: an orchestrator that computes nothingA DAG on a managed service has neither the memory nor the permissions for a training run. It triggers, it waits, it alerts somebody, and that last part is always written last, often after the incident.2022What happens around the codeTwo client engagements in parallel for much of the year. At that pace, the fatigue comes from reviewing, estimating and handing over, almost never from the model.2022Four feature flags, and how long they liveA classification that is not mine, and the question it leaves open: when the flag goes away. Without a written answer, it ends up as part of the product.2022Celery and FastAPI: answering before the work is doneA server that processes an image during the request eventually fails. Separating the acknowledgement from the processing changes the nature of the failure, and raises three questions the synchronous version concealed.2021A correct figure is not enoughJoining a consultancy, you learn quickly that the difficulty is less computing correctly than getting a figure accepted by people who did not follow the computation.2020Searching by resemblanceIn 2020, comparing documents by their vectors meant assembling the pipeline yourself. The state of the art at the time, and what vector databases have changed since.2020Two ways to open a nutThe chisel or patience: Grothendieck describes two ways of approaching a problem, with different costs. The choice depends on the situation more than on temperament.2020What lockdown left time to readA data warehouse, an economist and a psychologist. The year the time freed by not commuting went into books.2020What wears a developer down, and what holds them upWhat draws people into this trade is well documented. What drives them away is less so, and it is mistaken for lost interest. Three criteria account for most departures.2020Public figures do not make proofFrench open data is plentiful and free, which is what makes it misleading. What data journalism teaches an engineer, and the step almost nobody takes.2020What tidying does not settleThree method books read in the same year, and their shared limit: they organise a stock, while the scarce resource is attention. What remains of them once that promise is set aside.2020Negotiating scope, not priceFour traditions say the same thing about cooperation: the room to negotiate lies in what gets done, seldom in the amount. The position I took from them, and what it costs.2020A model that runs on the reader’s machineIn 2020 I ran recognition models inside a browser, for a reason unrelated to performance. A dated retrospective: the technique has been replaced, the argument still holds.2020A trade that found out it was a craftSoftware engineering started describing itself in workshop terms (code smells, refactoring, patterns) without deciding to. What that vocabulary lets you say about the trade, and its limits.2020Ten years before the trade had a nameFrom soil mechanics to data warehouses, a reading log kept since 2011 covers a decade in which the discipline was named only after the fact. What the log shows, and what it leaves out.
