Technology
From new silicon to a secure, AI-capable product
Most embedded products ship on a chip vendor's board support package:
often a forked kernel that is rarely upstreamed or independently
audited, and is frozen on an old release. We replace as much of it as
we can with software we build, audit and maintain ourselves, using AI
agents directed by our engineers.
AI-driven bring-up
AI agents write most of the board support package, and the hardware judges
the result: changes are tested on physical boards in our lab, with
remote power, serial consoles and scripted boot.
- Bootloader and secure firmware stages
- Linux kernel based on mainline (the official, publicly maintained kernel), with device trees and drivers
- Graphics, display, video and USB enablement
- New platforms on our own boot chain instead of the vendor's BSP
Firmware supply-chain security
Vendor firmware and binaries are inspected rather than trusted, and
findings are recorded as evidence, not asserted.
- Independent audit of vendor boot and secure-world (TrustZone) firmware
- Search for undocumented interfaces and open debug paths
- Agent review of our own firmware changes for input-validation and privilege flaws
- Working toward an auditable boot chain on maintained, open software
Our AI engine
One engine targets the board's CPUs, DSPs and NPUs without the chip
vendor's AI SDK.
- Reasoning, audio and vision models
- Backends for CPUs, DSPs and NPUs
- On-device inference, with no cloud round trip
Research: gaema, our
research arm, works on machine-learning systems, models and tools.
gaema.ai →