Who
My name is Matteo. I write about LLM systems as real technical objects: models, tools, context, memory, orchestration, costs, methods and limits.
My trade is systems and integration engineering, in a regulated banking environment, without interruption since August 2018, that is 8 years. I started as the field relay for the group's engineers on an enterprise telephony rollout, in a French private bank, from August 2018 to November 2020. Since then I have worked for the infrastructure arm of a French banking group: three years as a contractor, then a permanent hire in October 2023, in the same role. There I am the technical owner and level 3 authority of a multi-tenant provisioning ecosystem.
I name no employer on this site, and I publish no internal scope figures here. What gets measured here is measured on my own machine. The code is in my engineering portfolio, the rest on my LinkedIn profile.
What this site is
This site is a public, sourced corpus on large language models. The method is explicit: every statement is tied to a source, hypotheses are kept separate from facts, and negative results are archived alongside positive ones. The corpus currently holds 149 articles, each mirrored in English.
What I do
In parallel I run a personal LLM R&D lab where I instrument and measure multi-agent systems: orchestration, context management, memory, costs, tooling. The site publishes the stabilised results of that work, and draws more broadly on structured knowledge systems.
Since March 2026 I have built eight software products on my own in that setting. The largest, Ariane, is a retrieval-augmented enterprise document platform: 284,347 lines of Python as of 10 September 2026. Its search engine has its own measurement bench, 732 test functions, eighteen engine configurations frozen in a reference lock sealed with SHA-256, and a numeric regression policy rather than an impression (ndcg@10, a 0.01 margin, a paired t-test, a 0.05 threshold). Relevance is judged there on 5,572 TREC-format judgements, each with its provenance file. The thirteen architecture decisions are numbered and dated, and the production image runs unprivileged, on 1,311 pinned SHA-256 digests.
The limits matter just as much. That containerisation is local, on Docker: I have no experience of Kubernetes, of Helm, or of the public cloud. The measurement bench has neither continuous integration nor an orchestrator, it is launched by hand.
What I measure, and what I relay
I instrument multi-agent systems on my own machine: orchestration, context management, memory, cost, tooling. There, I measure, and the figures come from my own runs — 12 articles on this site are in that case, and each says so on its page.
On everything else, I read and I relay. I do not train models: what I write about training, hardware or internal architecture comes from the literature and from vendor documentation, not from experience. I am neither a lawyer nor an economist: when I write about the AI Act or a valuation, I read the text and cite the source — I do not interpret it professionally.
121 of the 149 articles on this site are in that second case. It is neither good nor bad: it is the real shape of what you are reading, and knowing where I stop tells you what to do with the rest.
Transparency
This site is produced with AI assistance. That is acknowledged and documented. See Made By AI, So What?
Contact
Direct contact: [email protected]. To report a factual error or a missing source on an article, see the dedicated procedure on the Method page.