The networks running on our engines don't come off the shelf from a framework: the architecture is ours, and so is the compiler that turns it into pure C. Both designed here, in Italy.
Edge AI is a stack. Silicon, the hardware around it, the software that runs on top, and the final application. We sit in the middle, where it is decided how small and how accurate a model can get: the network architecture and the compiler. In Italy, few design this layer rather than adopting someone else's.
We don't shrink general-purpose models until they fit a small chip. The architectures are built from the start to run on little memory and little current, and we design them ourselves.
A trained model comes out as pure C, with no framework and nothing to install on the device. The chain from data to binary is entirely ours.
Tell it which chip you have to run on and it searches for the engine and configuration that fit, instead of leaving the trial and error to you.
Every figure we publish comes from a public dataset with the split stated, so anyone can rerun the test and push back on us.
The methodology, the splits and the limits of each measurement are written out in full here: go to the benchmarks
Luviner is an Italian company, VAT IT02880910340, operating under EU law. No offshore entity between you and the people building the product.
Jessica Ferrari (CEO) and Filippo Groppi (CTO). Whoever answers your technical questions is whoever wrote the engine, not a first support tier.
Documentation, interface and support in both languages. Talk to us in either and ship your product wherever you need to.
Starting from a general-purpose framework means cutting until it fits. Designing the architecture lets us work backwards: start from the memory budget and build inside it.
What lands on the device is pure C. You don't inherit the lifecycle, the licensing or the change of direction of another company's framework.
Inference happens on the device. No traffic to a cloud, which makes conversations about GDPR and air-gapped industrial networks considerably shorter.
Upload a CSV of your signals and see what comes out. You can also just browse the engines.