What we work on.
Products cover a good share of the budget, and we compete for grants like any other lab. We still pick the problems ourselves. Two fields so far.
Language and voice.
We started here because the languages of this region get overlooked. Most models learn from English and are stretched to cover the rest, and it shows — names come out wrong, tone goes flat, the small things that carry a conversation slip through. We were sure there was room to do better, and there was. Canto and Tonic came out of it.
Knowledgebase systems.
Our second field, and we chose it for one reason: it is what we are best at. Orlo retrieves from your own documents more accurately than anything we have measured it against, and we keep the receipts. Expertise, turned into a product.
Annual plans can be refunded pro-rata within 30 days of renewal.1 After that window, the remaining balance converts to account credit for the next cycle. Refunds return to the original payment method within 5–7 business days.2
A Formal and Empirical Study on the Geometry of Multi-Hop Query Identification
Our first published paper. It asks when you can tell, from geometry alone, that a question needs more than one document to answer — the same work behind the boundary-hypothesis write-up below.
From the lab
Notes and write-ups from our research.
There is more coming.
We will share what we learn here as we go. If you want to work with us or just follow along, say hello.
Get in touch