Blog

Device identity for a fleet of cheap Linux boards around an Infineon OPTIGA Trust M: a key born in the chip, a five-minute JWT, enrolment by adding a public key, and where the chip's guarantee stops.

Doing edge AI takes two things that have nothing to do with each other: compressing the model so it fits a chip you can afford, and securing it so it stays your IP. Eighteen acquisitions in five years took almost every company doing the first. The second is still standing, and from September 2026 the CE marking asks about both.

What TrustZone does, how it works, and what Arm's Confidential Compute Architecture adds on top of it. With a note on when the hardware arrives.

The same fire watch job, built twice: YOLOv5s on a $249 Jetson at 10.2 W, and a 134k-parameter model on a €27 microcontroller in 340 ms. What the small board gives up, and what it took to get there.

Edge AI is not one kind of device. It runs on chips costing a dollar and on servers drawing a kilowatt. What changes across that range, and what does not.

Where post-quantum cryptography comes from, why edge devices need to be ready before the algorithms are mandatory, and what we will have to build when cosign ships PQC signatures.

PyTorch runs your model one layer at a time. A compiler fuses it into a single engine that runs far faster on the same chip, before you change anything.

Two models on one $249 box, fully offline: a vision model reads the image, a language model writes it up. Image in, plain-language report out, nothing sent to the cloud.

A quick review of the AI Act and the Cyber Resilience Act, and the implications for device makers.

A model has to run somewhere. Own a $249 box or rent a cloud GPU: one bill or a forever bill, and a choice about where your data lives.

A practical tour of Sigstore's cosign and Rekor, key-pair signing, cloud KMS, keyless OIDC, the transparency log — and how we verify it all in 150 lines of freestanding C.

A trained model is a long list of numbers. How small you can make each one on a $249 Jetson, and what it costs.

How to ship confidential AI models to edge devices while maintaining security and compliance.