Skip to main content

ScanClear (produit Digi4)2026Custom AI: documents, checks and processing

ScanClear — Website audit in 11 categories, with AI running on our own servers

A free tool that scans a website in under a minute and delivers a report with a fix for every issue, not just a score.

ScanClear, our product (https://scan-clear.com), audits a website in 11 categories (performance, SEO, security, accessibility, content…) and produces a PDF report with fixes. The AI is a small model hosted on our own servers: no data is sent to an AI provider.

The challenge

A score is not enough

Free audit tools give a score and a list of defects in technical jargon. Site owners do not know what to fix first, or how. And tools that add AI send page content to an external provider, which raises a real question under Moroccan law 09-08.

We wanted an audit with three properties: free and without an account for a first try, a concrete fix for every issue, and no data sent to a third party.

The solution

Deterministic checks first, AI second

  • One collection, 11 analyses in parallel: the page is loaded once (Puppeteer), then analysed by Lighthouse, axe-core (WCAG 2.1 accessibility), technology and outdated-library detection, and our own rules. Each category has its weight in the overall score.
  • Fixes reviewed by a person: 310 recommendations were written once, then reviewed. This catalogue is the main source of the report.
  • A local, bounded AI: a small model (Qwen2.5, 1.5 billion parameters) runs on our server's CPU. It assesses the page's text and fills missing fixes. Its answer must follow a strict format; an out-of-range score or an invented issue code is rejected. Its content score is averaged with the rule-based score.
  • The report ships even if the AI fails: AI enrichment comes after the report, never before.

The results

  • 11 categories, 274 checks and 310 written fixes.
  • An audit takes about 50 seconds in the usual case.
  • No data sent to an AI provider: the model runs on our servers.
  • PDF and Markdown export, site tracking over time for accounts.

Try it at scan-clear.com. What we learned here feeds our AI checks and validation projects.

Frequently asked questions

Why a local model rather than a well-known AI service?

To send nothing to a third party and keep costs under control. A small model is enough when the task is narrow and its output is checked.

Does the AI decide the score?

No. The score comes mostly from rules; the AI counts for half of the content score only.

Have a similar project?

Contact us to discuss it.

Message a Digi4 manager on WhatsApp