01
Proposal
Claude SonnetA first model understands your need and proposes an architecture, justified decisions and work packages, supported by your description and the architecture references.
Method & artificial intelligence
The power of Claude.
An explicit method.
Every step has a precise role. You see what was proposed, what was examined and what remains to be confirmed; the software checks between the steps and you decide.
Both roles are held by two distinct models, actually in service.
02
Linuen Studio controls
Linuen Studio controlsThe software checks the format, the cited references and consistency. An invalid response is a failure stated plainly, never a silent partial result.
03
Critical review by a second model
Claude OpusA second model, different from the first, examines the need, the scope, the architecture, the security and the effort assumptions. Every finding carries its consequence and the proposed correction; zero findings remains possible.
04
Revision if needed
Claude SonnetThe first model responds to every finding: it corrects, explains its disagreement, or flags missing information. One revision per assessment only, never presented as reviewed by the second model.
05
Final checks
Linuen Studio controlsThe same checks are run again on the revised version, and the application recalculates the estimate.
06
Delivery
Linuen Studio controlsYou receive the estimate, the report and a frozen version, with the exact status of the review and its caveats; the architecture decisions are recorded in the MADR format, readable by any architect. An email lets you know your project is ready.
Review mode
Two different models from the same company: the one that proposes is never the one that critiques.
Configured roles
Proposal
Proposes the first version, then revises it after the review.
In service: Claude Sonnet, at Anthropic.
Critical review
Critiques the proposal. It is always a model different from the first; if it were not, the review would be refused rather than replaced.
In service: Claude Opus, at Anthropic.
Revision
Revises the proposal in response to the findings; usually the same model as the proposal.
In service: Claude Sonnet, at Anthropic.
Personal information does not leave.
Before anything is sent to an AI model, the software detects the personal information in the project, on our servers in Canada and without AI, and replaces it with tokens. The model works on "[PERSON-1]" and "[EMAIL-1]"; you read names and addresses.
01
Detected as you write
Every text you save — need, answers, imported documents — is screened as soon as it is stored. Detection is local, deterministic and offline.
02
You decide, occurrence by occurrence
In "Your need", each detected value is shown to you, truncated, with two choices: mask or keep. No analysis leaves while a value awaits your decision.
03
Masked when sent
At request time, and only then, each masked value becomes a token that is stable across the whole project. The mapping stays in the database, in Canada; a safety net rereads the request for emails, phone numbers, cards and SINs.
04
Unmasked on return, for you only
The model's answer is unmasked on receipt: your project reads in clear text in your workspace. A project shared with a partner or published as an example stays masked.
What is detected: email, payment card number, social insurance number, phone number, date of birth, postal code, person's name.
It is a safeguard, not a guarantee: a trade secret, a sensitive figure or a name slipped into an ordinary sentence cannot be recognized mechanically. Send only what your project requires; the delivered project says when information was masked.
An AI review informs a decision.
It does not constitute a certification or a professional validation: the product issues none, and every deliverable says so. A revision is not presented as having been reviewed a second time if that review has not taken place.
- Two models, from the same company or from different companies, bring a diversity of perspective, not statistical independence of errors or a certification.
- The review covers the proposal and the project as transmitted; a figure changed afterward is not “reviewed” until a new cycle has taken place.
- A deliverable is “AI-assisted” or “demonstration,” and it declares so; the product does not replace a person.
- Excerpts transmitted to providers remain subject to the project's rights and the announced processing conditions; nothing is transmitted before an operation is launched.