What changed

Until now Ruby-Horizon ran on our infrastructure, and running it on your own hardware was something we scoped case by case. From today it is a product. The full Heyno workspace, with Ruby-Horizon inside it, ships on one workstation that we install in your building, on your network.

It is the same model, with the same behaviour and the same approval boundaries, whether it runs with us or with you. Where it is installed changes who can see your data. It does not change what the assistant will and will not do on your behalf.

Why we built it this way

The organizations that most need an assistant are often the ones least able to send their records anywhere. A county attorney and a clinic have the same objection, and no certificate answers it. The only answer that survives a procurement review is that the data did not leave.

What runs inside

The model, the retrieval index, the workspace and the tool layer. Calls are answered, records searched, meetings transcribed and mail drafted on the machine, and nothing about a request, whether prompt, document, answer or log, crosses the boundary unless you configure it to.

Speech, reasoning and retrieval share the machine’s GPUs. Live audio is strictly preemptive: a caller speaking interrupts a report being written, and the report resumes from where it was. Nobody notices a report finishing a minute later; everybody notices a pause on a phone call.

Diagnostics you own

The machine keeps a full local record of what it did and why, readable by your own staff without an account with us. When something goes wrong, the people in the building can see it rather than filing a ticket into the dark.

What you give up

Some speed of iteration. New model versions arrive as signed artifacts on a schedule rather than continuously, and you evaluate a version against your own held-out set before you promote it. Features that depend on our infrastructure arrive later, or not at all. We would rather say that before a contract than after.

On our own evaluations the local configuration lands within a few points of the hosted one on task completion and on knowing when to stop and ask, and a little lower on how elegantly it writes a summary. That is the trade we chose. A summary that reads flatter is acceptable; an assistant that books the wrong appointment is not.

Who it is for

Regulated industries, government and the agencies around it, and companies whose own customer contracts forbid third-party processing. In every case the blocker was never the model’s capability. It was the network path, and the machine removes it.

How to get it

It comes in four sizes, from a workstation for a small office to a rack-class build for a large one. Every size runs the full workspace; the difference is how many people it serves and how much history it holds. The first year is priced with the machine included, an annual fee follows from the second year, and the machine stays in your building for as long as you use it.

Tell us how many people will use it and what it needs to connect to, and we will size it with you.

See Local AI