Hosn is an on-premises AI company in Muscat. It deploys AI at the customer's own site, for institutions in Oman that cannot use shared AI infrastructure. I co-founded Hosn AI Services LLC with Mohammed Al-Adawi.
Why we started it
Most AI products run on a provider's servers. For many users, that works well. For some institutions in Oman, it does not. Their data cannot leave the building. Their policy decides where each record lives.
At BHD Group, we use applied AI in our live operations. Hosn serves a different need. It is for institutions that cannot send their documents to a shared service.
I wrote about that need in a separate essay on AI that stays on-premises. This article is about the company we built to meet it.
A partnership
Mohammed Al-Adawi and I co-founded Hosn as equal partners. Hosn is its own company, Hosn AI Services LLC. It operates within the BHD Group portfolio, next to products such as Mithaq and Cardify.
What Hosn does
Hosn deploys AI at the customer's site. Before any work starts, we write three boundaries down with the institution.
The first is the site boundary. We review the location, network routes, storage, identity and administration with the institution. We record them in the delivery scope. An air-gapped option is available, with no internet connection at all. In that case, we also write down the update procedure, because nothing arrives by itself.
The second is the language boundary. We test Arabic and English behaviour on representative, approved material. We confirm the model and the acceptance criteria only after that test. Selected open model families, such as Gemma and Qwen, can be evaluated.
The third is the commercial boundary. The terms go into a written quotation and contract. Maintenance and update support are scoped separately.
Configurations by quotation
Hosn plans work from three starting points. Kernel is a single appliance for a department-level workload. Tower is a multi-node setup for several teams. Rack is rack-scale GPU infrastructure for institution-wide workloads.
Every configuration is priced by quotation. Hardware, integration, network boundary and support differ for each institution. A fixed public price would hide those differences. So the website publishes no prices.
Before a quotation, an institution can ask for a Sovereign AI Readiness Assessment. It documents the workload, the users, the approved documents, the integrations and the facility. Its output is a written scope. The institution's technical and procurement teams can review that scope before we prepare a quotation.
What the work looks like
The workloads are practical. A team asks Hosn to summarise an approved document set and link each result to its source. Another team asks for an Arabic draft from approved notes, with every missing fact marked for review. Others extract fields from forms, or search internal procedures.
Each of these tasks has one rule in common. The answer must point back to a source that a person can check.
What the website states
Hosn publishes no customer names. The website lists workload categories, such as public institutions, regulated enterprises and healthcare. These categories are not customer references.
The site also answers a common question: how long does deployment take? The schedule depends on procurement, hardware delivery, site readiness, integration scope and acceptance testing. We include a delivery plan with the quotation.
What I have learned so far
Hosn is a young company. One lesson stands out so far. The model is only one part of the work. Most of the work sits in the boundary. Where does the system run? Who administers it? How do we test Arabic? We write those points down first, and then we test the model.
You can read more about Hosn at hosn.om.
Related: Why some institutions in Oman need AI that stays on-premises, Ventures and leadership