Why it matters
The core principles
We focus on 3 main principles to help you integrate your AI model into your edge device.
Model optimization
The model is optimized to run on the device, with minimal latency and power consumption.
Confidentiality
The model stays encrypted on the device and is only decrypted inside the trusted execution environment.
Attested integrity
Unauthorized changes to the model or its runtime are detectable. Our tooling attests the model’s integrity, so that you can remain compliant.
What we do
Protection at rest and at runtime
The model is encrypted on disk and only decrypted inside the device’s trusted execution environment.
The weights are never exposed in the clear, so the model can’t be copied off the device, cloned, or swapped for a tampered one. Even by a malicious actor with root access, the weights are protected.
Model engineering taylored for the target
We can handle the model work end to end — training, quantization, and platform optimization — so it fits the latency, memory, and power budget of the chip you’re shipping, without giving up the accuracy your task needs.
Legal and compliance out of the box - EU AI Act and CRA
The EU AI Act — whose obligations for high-risk systems apply from August 2026 — requires protection against unauthorized model changes, and the Cyber Resilience Act (CRA) adds security obligations for connected products from December 2027.
Attestation and tamper-evidence are built in from the start, giving you the technical measures these new rules expect.
Secure and flexible key management
Our toolchain is compatible with Sigstore Cosign. We chose to integrate with this tool to offer flexibility with regards to the key management.
You can either use a physical HSM (Yubikey), or a hosted HSM from different cloud providers (Google Cloud, AWS, Azure).
Runs on hardware you already have
Integration is light: if your device is running with an ARM CPU, it essentially already works. No external secure element are required, we make the most of any hardware security the chip does offer.
We are already running on silicon from NVIDIA, MediaTek, NXP, Texas Instruments, and Rockchip.
Licensing that fits your product
We offer white-labeling, pay per device, or a custom license — whichever matches how you ship.
Product flexibility
Only interested in the model optimization, or the model protection? You can use either or both, depending on your needs.