It walks a company from commercial prospecting through to compliance with the European AI Act. Eleven agents, one case file per project, and an implementation plan that writes itself. All inside the house: local model, local index, zero cloud.
From message to plan
What it is
Getting AI into a company is a documentary problem before it is a technical one: you have to analyse the business, judge whether its data is good for anything, choose the approach, justify the investment, plan the rollout and answer for all of it under a European regulation. That work is done today with a chat open in one tab and a folder of loose documents, and it is lost in full between one client and the next. Lideria turns it into a system: each module produces an artefact, each artefact fills in the section of a twelve-part implementation plan that belongs to it, and everything stays in the project's memory — indexed and searchable by meaning — to feed the next question. And it runs entirely on the machine of whoever uses it.
Who it's for
Whoever puts AI into companies: consultancies, innovation departments and internal teams who have to leave a written record of what they did, why, and under which legal obligation — and who work with information that cannot leave the organisation.
How it works
A.R.I.A. decides before answering
The orchestrator does not answer straight away: it first asks the model whether the request is resolved by talking or whether it must delegate, and to whom. If it delegates, it runs the specialised agent and rewrites its output for the user without showing the kitchen. It lives in two places at once: the main chat and a floating panel that knows which screen you are on and proposes the missing action.
Eleven agents, five modules, seventeen workbenches
Commercial strategy, solution factory, governance and MCP Hub. Market → viability → proposal run chained: each step receives the previous one's context. And wherever there is a calculation that must not be invented — the viability score, the security posture — the backend does it with a fixed formula, not the model.
The implementation plan fills itself in
Twelve sections, from context preparation to annexes. Each kind of artefact knows which section it belongs to: a market scan lands in 2, a viability study in 3, a proposal in 4. The day's work becomes a document with nobody copying it across by hand, and the percentage of plan completed is visible from the first screen.
One case file per project, with a memory of its own
Each project has a file of its own: artefacts, activity, the plan and a semantic index built with local embeddings. When you ask something, the system retrieves by meaning what was already said in that project and puts it into context. The memory is not global or shared: it belongs to the case file.
MCP client written from scratch
JSON-RPC 2.0 over standard input and output, without using the SDK: handshake, tool listing and real execution against external servers. The explorer builds each tool's form from its own schema, so a new tool shows up usable without a line of interface being touched.
The AI Act as a map, not an opinion
Eighteen categories of use classified by risk level with their list of obligations, resolved by a deterministic table. The agent writes the opinion, but cannot move the level: that is decided by the regulation, not the model. It is the difference between a report that holds up in front of an inspector and one that sounds good.
Frequently asked
Why a small local model instead of a large one in the cloud?
Because sovereignty was the requirement, not decoration: a client's case file — what systems they have, what data, where they are weak — is sensitive material, and the platform was designed so that it never left the machine. Today its great strength is structuring and deciding; for long-form drafting, the next step is a gateway with profiles: routing and embeddings stay local, long-form drafting may go out to a frontier model without the client's identifiers, and a fully sovereign mode remains for anyone who requires it by contract. The architecture is already prepared for that, and it is the next piece of work.
It works end to end and it is tested. The next two steps towards opening it to clients: the model gateway, to draft long documents at a quality worth sending, and an interview engine that collects the information straight from the client.