CMX in 2026: what exists today
A fair question
We publish opinions here about architecture and AI. A reader is entitled to ask what we have actually built. This is the inventory in mid-November 2026.
One rule first. We do not name clients, and we do not describe their projects in a way their own teams would recognize. That rule protects them today and it will protect you.
A governed door for generative AI
Central IA Souveraine is our product for regulated institutions. Every AI use goes through one internal address, eleven controls run between the employee and the model provider, and each decision lands in a hash-chained ledger that an auditor can verify. It is deployed in the institution's own cluster. The console has 21 screens in French, Arabic and English.
It comes with a knowledge vault: each document exists once, inside the perimeter, and the model receives only the passages that answer the question.
Read the eleven controls.
A workflow engine that survives a crash
hwf is a C++20 prototype of a durable engine for processes that mix code, AI agents, human decisions and timers. Every state change goes to an append-only journal. We kill the engine in the middle of a loan booking and the loan is booked once.
Measured on an Apple M5 Max: 50 to 70 microseconds per durably committed step. It is a prototype, and the post lists what production still needs.
Read the engine, its numbers and its limits. The research line on the workflow model itself is in the 2yFlow technical report.
A method that attacks our own designs
Our method works down eight levels, from intent to software specification, and each step ends at a gate that a person passes. Before a design moves on, we try to break it three ways: explore its states, check its numbers, test its resilience.
We applied that to a product of our own, a tracker for trucks. Ten problems came out of a first design that looked serious.
Read the design journal.
Reading field documents without guessing
We built a service that reads paper documents photographed on a phone, on the operator's own server. Three OCR models read every field. A value is accepted only when the evidence agrees: the readers are unanimous, the arithmetic closes, the names match known reference data. Otherwise the document goes to a person, with the reasons.
The production service is written in C++. The phone app keeps each photo until the server has confirmed it, including through a crash or a dead network.
Our first target for automatic acceptance was missed. Too many documents still go to review. We prefer that failure to the other one, which is a wrong value stored as accepted.
Twelve chairs
Our 2027 final-year internships are four squads of three chairs, from February to June in Casablanca. Selection ends with a Talent Acquisition Day in January 2027. The rules are public: the Squads Game.
What is still missing
The workflow engine is a prototype. The tracker is a set of proposals and no truck has been tested. And we have published no client reference. We will when a client agrees in writing. Not before.