Book a demo

Powerful Enterprise AI. Private by design.

The AI appliance for sensitive industries and governments whose data can't leave the building.

Book a demo How it works

Product footage shows a fictional demo company.A simplified view of Ouroboros OS, drawn for this page.

Employees ask in Arabic or English. Answers about the company come only from documents they're cleared to see, and cite their source. Nothing leaves the building without a person's approval.

Explore the product

Every answer cites its source.

Ouroboros OS states company facts only from documents the employee is allowed to read. Hover or tap a citation to see the document, the page and the passage it came from.

  • Answers in Arabic or English, in a right-to-left Arabic interface. Ask in Arabic about an English document and the answer comes back in Arabic, cited.
  • Reads scanned documents and spreadsheets.
  • If the documents an employee may read don't hold the answer, it refuses with one fixed sentence, “Insufficient authorized context.”, instead of guessing, and tells the employee why.

What are the site induction rules for contractors?

Ouroboros Assist Contractors complete site induction before their first shift. The certificate stays valid for twelve months.
1HSE ManualPage 4
All contractors complete site induction before their first shift on any site.
2Training PolicyPage 2
Induction certificates remain valid for twelve months from the date of issue.
Not enough evidence to answerAnswers come only from documents you may see.When the answer isn't there. The fixed reply: “Insufficient authorized context.”

Each employee sees only what they may.

Permissions are enforced in retrieval, before the model sees anything. The same question from two employees gets two different answers, depending on clearance.

  • Single sign-on through authentik, built to federate with Active Directory, LDAP or Entra.
  • Users and their departments kept in sync from authentik with SCIM.
  • Every answer is written to the audit ledger.

What capital budget did the board approve for next year?

FL
Finance leadRestricted clearance
Permission filter

The board approved next year's capital plan, with the largest share for grid upgrades1.

SE
Site engineerStandard clearance
Permission filter

Not enough evidence to answerAnswers come only from documents you may see.

Nothing leaves without approval.

Assist does multi-step work across mail, calendar and Jira. Before anything leaves the deployment, it stops and shows exactly what will be sent.

  • Decline, and the step is skipped. Nothing is sent.
  • Every decision is recorded in the audit log.

Reply to the supplier with the new delivery dates.

Read an email messageDone
List calendar eventsDone
Create an email draftDone
Send an emailApproval requiredDoneDenied · Approval declined
Approval required

“Send an email” needs your approval before it runs.

This will leave your deployment:

To
Supplier contact
Subject
Revised delivery dates

Approved and sent. Recorded in the audit log.

Approval declined. Nothing was sent. Recorded in the audit log.

Every answer carries a signed receipt.

Each answer is signed and written to a ledger. An auditor can re-check a receipt later, long after the conversation is over.

  • The answer text matches the stored record.
  • The signature is valid.
  • The ledger entry is found.

Your security team can verify the receipts themselves.

Signed answer

Answer
Complete
Signed
31 Oct 2026, 14:02
Receipt ID
3f9ac47e0b12
  • Answer text matches
  • Signature valid
  • Ledger entry found

Not checked yetChecking…Matches the stored record · checked 14:02

One governed path to every answer.

Every request follows the same path, and all of it runs inside the customer's building.

Your building
  1. EmployeesChat, Assist and Admin, white-labelled, in Arabic and English.
  2. Identityauthentik single sign-on, built to federate with Active Directory, LDAP or Entra, with SCIM user sync.
  3. Harness ContractOur versioned rulebook for every request: citations required, the fixed refusal sentence, prompt-injection screening, approval gates and signed receipts.
  4. RetrievalPermission-filtered search over a Qdrant vector store with bge-m3 embeddings, with OCR for scanned documents. A knowledge graph is built alongside; it is not yet used in answers.
  5. AgentsMail, calendar and Jira connectors, tested on demo servers so far, with governed skills and memory.
  6. ModelsOpen models, including our own Sayqal models, served locally with llama.cpp (vLLM planned for the 128 GB box). No cloud API calls.
  7. HardwareDesigned for one 128 GB box (NVIDIA DGX Spark class) or a 96 GB RTX PRO 6000 workstation, or a GPU server for larger sites.
  8. Audit ledgerEvery answer, approval and admin change.
Security and sovereignty

It learns. People decide what it keeps.

After a task, Assist saves what it learned as a reusable skill, which the employee can retire, and asks before it remembers a fact. Skills and memories are scoped by permission, reviewed and audited.

Skill savedAssist saved a skill: weekly-site-safety-summary
Retire
Remember this?Assist suggests remembering:
Our fiscal year starts in April.
DismissRemember

Reviewed, and recorded in the audit log.

Their brand. Their language.

Customers put their own name, logo and colours on the sign-in page and every screen from the admin Branding panel; every employee sees the change on their next page load. The “Powered by Ouroboros OS” credit is optional.

Own your intelligence.

Sayqal is our own model line: open models, each tuned for one job in the product, such as cited answers, tool calls and injection screening, in Arabic and English. They run on the customer's box, and the customer keeps them.

Meet the Sayqal line

Sayqal Chat
98.3%on our 292-item internal English evaluation, against 87.3% for the base model.
Sayqal Chat98.3%
Base model87.3%
Tool calls17/17
Base model12/17
Sayqal Guard
22–24of 25 attacks per set caught, on a blind test written before training, with 1–3 false alarms.

Screens prompt injection in English and Arabic.

Attacks caught22–24
Caught in every setCaught in some setsMissed

Both models are live in the MVP. These are internal evaluations. The chat evaluation is in English and in-distribution; there is no Arabic chat score yet. A test on real documents is next.

Your language, from the ground up.

We built Ouroboros OS and Sayqal for Arabic and English. Arabic is the hard case: if Sayqal can do Arabic, it can do the next language. For Swedish, Polish or Malay the method is the same, with data and a test set in that language and a short tune on an open model. No third language is done yet.

كل إجابة تستشهد بمصدرها.

“Every answer cites its source,” in Arabic. It is written right to left, its letters change shape with their place in a word, most of its words grow from three-letter roots, and the dialects people speak differ from the standard they write.

Tuned for each language

Sayqal is tuned for Arabic and English today. A new language gets its own data and test set and a short tune on an open model, and the tuned models stay on your box.

Tested in each language

We test models in the languages they answer in, and say which figures are internal and what isn't measured yet. Our injection guard, page search and graph links have Arabic blind-test scores; our chat model doesn't yet.

A product that reads your way

The interface, writing direction, refusal notice and sign-in page, in Arabic or English and in your brand.

Built for critical organisations.

Sensitive industries and governments whose data can't leave the building, in any country.

  1. Banks and financial services

    Customer and transaction data that regulators expect to stay in the country.

  2. Hospitals and healthcare

    Patient records and clinical documents that often can't be sent to a cloud model.

  3. Governments and the public sector

    Ministries, agencies and their suppliers, bound by strict data-handling rules.

  4. Energy and critical infrastructure

    Operating procedures, drawings and safety documents for national assets.

Why AI belongs inside the walls.

Employees already use AI at work. The question is where their data goes.

47%

of enterprise AI conversations run through personal accounts, outside company security. More than 6% of all enterprise AI conversations contain sensitive data.

LayerX, State of AI Usage Report 2026 (May 2026)

92%

of organisations with an AI-related breach lacked proper AI access controls.

IBM, Cost of a Data Breach 2026

~4months

is roughly how far open-weight models trail the closed frontier, on average.

Epoch AI, May 2026

~40%

of AI's value rests on sovereign or “sovereign-enough” solutions.

McKinsey, Mar 2026

Bring AI into your building.

See Ouroboros OS answer, act and sign its work on a fictional demo company's documents. We're looking for our first pilot partners.

Book a demo Talk to us
Where we are

A working MVP that runs end to end on a development box, with a fictional 457-document demo company and six employee personas. No customers or pilots yet.