Super Intelligence
The top layer is where the reasoning happens. Hurukan draws on several models: each task — a draft, an analysis, a classification, a change to the system itself — goes to the model that fits it, within rules the company sets about quality, cost and where its data may go.
- Frontier models and local models behind one interface
- Each task routed by quality, cost and data sensitivity
- Rules about which data may reach which model
- Microsoft 365 Copilot, Claude and ChatGPT integrations planned
What the layer is for
Most of the work in a company that needs judgement also needs reading: an email from a customer, a supplier’s offer, a contract clause, a set of numbers that do not agree. Superintelligence is the word we use for the capability to do that reading and reasoning at the level of the most capable models available. This layer is where Hurukan reaches it.
It is deliberately a set of models. The model best at summarising a complaint differs from the one best at checking an invoice, and if the company has decided that personnel files stay on its own hardware, only models running there will see them. So the layer is a router with rules rather than a single connection.
How a task is routed
A request arrives from an agent with three facts attached: what the task is, which records it touches, and how those records are classified. The router checks the company’s rules — this class of data may only go to local models; this kind of task needs a frontier model; this one should be cheap — and picks a model that satisfies all of them. The choice, the cost and the result are logged against the task.
If none of the allowed models is good enough for the task, the request comes back to a person, so data reaches only the models the rules permit.
Above and below
Above this layer are only the people who use the system. Below it, Agentic Logic is its only caller: agents ask for reasoning, and the answer returns to them rather than straight into the record. Agents write to the record within their limits; models supply the reasoning. What a model may see is decided by the Ontology, which says what each record is, and by Security and Sovereignty, which says who is asking and what they may read. That ordering is the point: the most capable component is the one with the least direct power.
What is planned
Integrations with Microsoft 365 Copilot, Claude and ChatGPT are planned, as is support for local open-weight models on your own hardware. None of them is live yet, and this page will be updated as each one becomes available.
What the layer does
Routing per task
Every request carries its purpose and the classification of its data. The router picks a model that is allowed to see that data and is good enough for the job, and records the choice.
Frontier models
Where a task benefits from the most capable models available, it can go to a frontier provider through the company's own account. Planned for Claude, ChatGPT and Microsoft 365 Copilot.
Local models
Open-weight models running on hardware you control, for work whose data stays there: personnel records, contracts, anything you mark as such.
Grounded in the record
Models answer from the company's own data model and documents, with the sources attached, rather than from general knowledge alone.
Usage you can see
Every call is logged: which model, for which task, on whose behalf and at what cost, so every cost can be traced to a task and the person it served.
Replaceable by design
Models change every few months. The layer treats them as interchangeable parts, so one can be replaced while the agents above it and the ontology below stay as they are.
Works with
Decisions and Tasks
One approval inbox for agent proposals with sources and risk, and tasks that always have an owner.
Customer Service
Cases with stages, SLAs and escalation, the customer's full history, and agents drafting answers for people to send.
Marketing
B2B funnel, target accounts, campaigns and content, with lead scoring that shows its evidence.
Questions
Is superintelligence a claim that the system thinks for itself?
It names the capability the layer is built to use: the most capable models available, applied to the company's own record. Judgement on anything critical stays with people, and the layers below enforce that.
Which models will it support?
The design covers Microsoft 365 Copilot, Claude, ChatGPT and local open-weight models, with the routing layer choosing between them for each task. The named integrations are planned and not live yet.
Will our data be used to train someone else's model?
You decide which data may reach which provider, and sensitive work can stay on local models that run on your own infrastructure. What a frontier provider may do with requests is governed by your agreement with that provider, which is exactly why the routing rules are yours to set.
Build the company's operating system with us
Hurukan is pre-launch. If you run a company and want a sovereign, agentic system rather than another suite, we would like to hear what you need.