A neural network on a microcontroller.
The foundation rests on a single idea: learning must happen on the machine, from a measurable physical quantity, without ever leaving the machine. No generative model.
Six building blocks.
NNUE
Efficiently Updatable Neural Network: a network designed for incremental update. Everything is not recomputed at every cycle — only what has changed is corrected.
origin: chess engine · topology of the machine version: not publishedRust no_std
Memory safety checked at compile time: the code is written in safe Rust, apart from a few isolated and commented hardware register reads. No garbage collector, no hidden runtime.
target ARM Cortex-M (STM32) · Xtensa LX7 (ESP32)Local self-calibration
The system is designed to learn on the machine itself, from a measurable physical quantity, without ever leaving the machine. The exact mechanism is not published. Validated in simulation; the first real-machine trial is still ahead.
local learning · mechanism not disclosedLearning without labels
No human annotation is required: it is a physical measurement, read on the machine, that guides the learning. The metric, its form and its coefficients are not published.
no human label · metric not publishedOffline by design
No network connection by default. The design plans to pot the sensitive components in resin and to erase the keys on any opening.
chosen design: Secure Boot · encrypted flash · tamper detection — not yet implementedSovereign microcontrollers
Traceable and available components, with no dependence on a single supplier.
STM32 · ESP32 · RISC-V (exploratory)Self-calibration, without the detail.
The module's self-calibration method is a research program specific to the laboratory. It is not published: neither the mechanism, nor the metric, nor the network topology. It belongs to the laboratory's know-how (trade secret) and is disclosed, where applicable, only under a confidentiality agreement. It is not in the module offered today, which observes without ever controlling; it will only come out after a real measurement.
The five protections.
This is the design chosen for the module, layer by layer. None is industrialized yet.
| Layer | Implementation | Effect |
|---|---|---|
| 1. Potting | Two-component epoxy resin, full encapsulation | Forced opening = destruction |
| 2. Tamper detection | Mechanical switch under the potting + track mesh | Erasing of the keys in RAM |
| 3. Secure Boot | Mandatory signed firmware, encrypted flash | Any unsigned binary is refused |
| 4. Silicon locking | Read-out protection level 2 + dedicated secure element | JTAG/SWD access permanently blocked |
| 5. No delivery of the source | No source code, no final weights transmitted | The client receives a binary, not know-how |
The portfolio and the technology are two separate things. What gets installed ships as a signed binary. What gets licensed is a documentary dossier. Never the other way around.
Any machine that consumes energy and wears out.
Batteries & inverters
Calibration of the discharge signature, early detection of a failing cell.
Fans & pumps
Detection of imbalance, cavitation, excess torque. Maintenance before breakdown.
Electric assistance
Reading of real use (terrain, load) to estimate range, with no data transmission.
Refurbished goods
Learning the wear of a second-hand device, extending its service life.
Positioned as a complement, never as a substitute. The module is designed to be added to an existing machine. It is intended to observe and warn, without taking control of it.