One person, two systems, published numbers.
Latency-critical machine-learning systems that run locally, on hardware the operator controls.
ayaka is the trade name of a sole proprietorship in Bayreuth, Germany, registered on 11 August 2026 at the Gewerbeamt Bayreuth. The owner is Luca Finn Law, a self-taught engineer, and the business consists of that one person. There is no team, no funding round and no office — which is worth stating plainly, because a site of this kind usually implies otherwise.
What local-first means here
Not an ideology. A constraint that two very different problems turned out to share.
A prosthetic grip has to be decided before the hand arrives at the object, inside about 125 milliseconds. That budget rules out a network round trip, so the model that decides has to sit on the device. A translation pipeline reading an entire book has a different reason for the same answer: that text belongs to whoever wrote it, and shipping it to a rented endpoint is a decision nobody asked for.
In practice: models are trained on hardware here, they are exported to formats that run without a vendor, and the figures describing them are measured on the machines they actually run on.
Working method
Publish the numbers, including the bad ones.
The Hime page says that the fine-tuned models do not currently beat their own base model, and that an earlier positive result was withdrawn after the evaluation harness was found to be leaking its answer key. The ayaka-hand page says that no physical hand exists and that the hazard flags are wired to nothing.
Both of those cost something to say and are worth more than the alternative. A result that has never been allowed to fail has not been tested.
Registered, but not offered as packaged work
The trade registration also covers game development, and research and development in software and machine learning including entry into research competitions. Those are real fields of activity rather than services on a list, so they are named here instead of on the services page.