Applied AI lab · Reno, Nevada
Words that last.
AI talks a lot and finishes little. Orkhon builds models and agents that do real work in real businesses.
Routine decisions go to the fast model. Uncertain ones go up.
A System One model is fast and narrow: it returns a typed decision with a probability instead of prose. Routine, narrow decisions go to it. When confidence is low, the decision escalates to a reasoning model or a human.
How System One models workEvery decision carries its probability and an audit trail, in either lane.
Bengü-1
Open model · MIT · on Hugging Face
bengü
Old Turkic: eternal. The Orkhon stones call themselves bengü taş, the eternal stone.
Our first model sorts incoming mail. For each mail it gives three answers: a category, whether the mail is protected, and whether it needs a reply. It runs on the CPU, and the mail never leaves the device.
- Input
- One incoming mail
- Output
- A category, protected, needs reply
- Writes text
- No
- Result
- 0.963 group macro-F1 on the public triage track
- Status
- Open model, MIT, on Hugging Face
- Base models
- bge-m3 and xlm-roberta-large, MIT
What we build
Usta takes on the daily work of running a restaurant, starting in our own kitchen. Ermut is a set of free character tests on how you meet people and work with them.
Lab plus operator
Usta runs first in our own kitchen: Piatto Reno, Reno, Nevada.
Bora Mediterranean follows when it opens. Our products run first in businesses we operate ourselves. Results at the register come before benchmarks.
About the lab
The name
The Orkhon inscriptions were carved in the Orkhon Valley of Mongolia in the 730s. They are among the oldest surviving texts in any Turkic language, and they call themselves bengü taş: the eternal stone. After 1,300 years they are still standing.
The name and the mark