Nothing leaves your network. Your engineers decide everything. Requirement matrices, execution-document review, site reports: the tool reads, cross-checks and flags — it never signs off in their place.
Module under construction · pilot programme open
# on a Linux server inside your network $ docker compose up -d ✔ pool + database started — everything is in the image # open https://<server>:8443 → first-boot wizard # then pair your PCs and GPU servers $ enable « Engineering » in the admin dashboard ✔ 0 outbound connection
Tender files and specifications — and one missed requirement discovered after signature.
Design notes, data sheets and shop drawings to approve within contractual deadlines.
Meetings and site visits written up after the day is over.
Water, energy, industrial and healthcare clients forbid sending their files to an online AI service.
Mobility, water, energy, buildings, industry: each domain brings its own vocabulary, document types and checklists, set by your administrator — no development needed.
| For | Function |
|---|---|
| Studies, tenders | Requirement matrix: every requirement with its number, page and close passages from your own references; vague requirements flagged with a draft question to the client. |
| Project management | Execution-document review: for each submitted document, the relevant requirements of the work package, the values compared, the review register and its deadlines. |
| Site supervision | Dictated report on site, laid out for proofreading. |
| Tenders | Inconsistencies between documents: a specification clause with no line in the bill of quantities, different units. |
| Owner's engineer, turnkey | Brief compared with design; main-contract requirements flowed down to each subcontractor. |
Why an engineering firm needs an AI that stays in-house, how the platform is installed, and what the module will do. Animated illustration: the module is under construction.
Documents are read without being kept, only by the machines your administrator has classed as trusted servers. The engineer's work — positions, comments, reviews — stays encrypted on their own device: neither management nor the administrator reads it.
The central server — the pool — ships as a single Docker image containing everything: AI, document search, admin dashboard, employee space, PDF / Word rendering, and distribution of models and software to your machines. No external service, no dependency to install. It arrives blank: a first-boot wizard creates the administrator, the HTTPS certificate and the settings.
A small program deployed on workstations (Windows, Linux) with your usual tools (Intune, GPO, Ansible). It uses the processor and graphics card while the PC is lightly loaded. The pool sends it the model suited to its hardware and its updates; it keeps no data.
Why it matters: computing capacity grows with your fleet, without buying a server farm. Each PC is paired by the administrator.
The same kind of program, placed in front of a GPU server running a large model. It connects it to the pool and can fail over to another pool if its own goes down.
Why it matters: large models and confidential jobs run on machines the administrator has classed as “trusted server”, never on a workstation.
One pool per site. Pools recognise each other by signature and copy what you choose to replicate (memory, model catalogue, settings), with no master pool: any of them can take over.
Why it matters: losing a server or a site does not lose the firm's knowledge, and each site keeps its data at home. Reserved projects are never copied without your consent.
Every night, a small learning complement (LoRA) is trained on your servers from what is new in the knowledge bases opened for this purpose: your vocabulary, your methods, your references. It is versioned, evaluated before activation, and can be rolled back. Confidential projects, reserved bases and the data of people who object are excluded by design. Nothing leaves your servers.
Why it matters: the firm's know-how no longer leaves with the engineer at the end of a mission — it becomes part of the AI everyone uses.
The pool speaks the same language as the OpenAI API: your engineers plug in their terminal or editor tools with a sk-cellule-… key issued by the administrator. Calculation scripts, macros, in-house tools: the coding assistant runs on your AI.
# OpenCode: an “OpenAI-compatible” provider
baseURL = https://<pool>:8443/v1
apiKey = sk-cellule-…137 settings in 19 categories, each explained in the admin dashboard: every option is a switch your IT department controls, with safe defaults, applied without a restart, and switched back at any time. Animated illustration generated from the product's actual settings registry.
Every requirement is listed, numbered and traced to its page. Gaps in numbering and vague wording become questions — before signature, not after.
Less time cutting and pasting specifications, more time on technical judgement. Reports are dictated on site instead of typed at night.
An AI you can use on files your clients will never let into the cloud — with a level of confidentiality per project and separation between competing clients.
Existing PCs and GPU servers do the work. No per-token billing, no data sent to a foreign provider, no dependency on an internet connection.
References, lessons learned and methods live in your knowledge bases, replicated across your sites — the firm's memory survives turnover.
One pool per office or country, data kept in its jurisdiction, administration centralised. Start with one office, extend through federation.
On your server, half a day; settings with your IT department.
A complete tender file as a matrix, the reviews of a work package, site reports.
Trusted servers, no outbound traffic, separation between projects.
A qualification protocol (e.g. pharma), no requirement lost.
Requirements missed compared with a matrix done by hand on the same file (target: zero) · time per file and per review, before and after · inconsistencies found and confirmed · no outbound traffic throughout, recorded by your IT department. Measurements are entered by your teams: the tool collects no usage statistics.
A volunteer team of 3 to 10 engineers and a referent · a server in your network (sizing provided beforehand) · two or three real or archived files per domain, chosen by you · a 30-minute check-in every two weeks.
Pilot places are limited. Tell us about your firm, your domains and the files that keep your teams up at night.