Fast
Our models are small, so they compute in seconds or minutes, not hours. Fast enough to re-plan whenever reality changes.
Knowledge instead of data.
We build very small models that solve hard industrial problems in minutes, on ordinary hardware, with little data, poor-quality data, or none at all.
Our models are small, so they compute in seconds or minutes, not hours. Fast enough to re-plan whenever reality changes.
A laptop or a small embedded board is enough. No GPU cluster, no supercomputer, no cloud bill.
No large data-collection project before anything works. Our models start from what is already known about your process.
Blurry images, gaps, sensors that fail: our models keep working, because they don't depend on the data alone.
There is no magic here. Large AI models learn everything from examples, which is why they need so much data and computing power.
We take a different route. We build into the model what is already known: the rules of your business, the constraints of your process, and generally accepted engineering knowledge. The model then only has to learn, or search for, what is truly unknown. This is what our Know-How-Augmented AI and Live Models are made of.
The result is a model that is small, fast, and robust, built for one problem: yours.
For years, the dominant direction in AI has been bigger models, more data, and more computing power.
We have argued the opposite for a long time, well before it was fashionable: that for real industrial problems, models can be far smaller than people think. Today, more and more of the field is moving in this direction. We have been applying it in industry for years.
Our approach won an international optimization challenge organized by the Aisin Group, of the Toyota group, against more than 800 AI companies. It led to a multi-year engagement in which we optimized a full supply chain, with 30% fewer trucks and a cost reduction of 30%, both validated by the client. The client’s engineers had expected the problem to need about 12 hours on high-performance computing. Our model solved it in about 4 minutes on a laptop. Read the case study.
An industrial client needed real-time detection of small, nearly identical objects, some of which cannot be told apart by the human eye. The images were blurry, and the camera regularly shut down from the heat, leaving the system with no images at all. We built a very small model that combines whatever images are available with knowledge of the process. It runs in real time on a Jetson Nano, a compact, low-cost embedded computer, and was accepted by the client after testing. More on real-time computer vision.
Because each of our models is built for one client’s process, usually under confidentiality. There is no generic model to benchmark, and our clients’ problems are not ours to publish. We are happy to show results in more detail under NDA.
If you work on small models, knowledge-based AI, or optimization, and want to exchange ideas, we would be glad to hear from you.
Aisin first worked with Funartech in 2021. Funartech provided us with a solution that met our expectations based on their hard work.
We asked Funartech for a more advanced experimental project in 2022. We were looking to improve logistics from a long-term perspective. This project was an important and difficult problem to solve. They were willing to take on the challenge and successfully developed a prototype. We are pleased with their customer support and look forward to continuing to work with them.
Tell us about it. We'll tell you honestly whether a small model can do the job.
Contact us