AI Report Reader
AI reads vendor maintenance reports and proposes work orders; a human validates.
- Role
- Review screens, demo mode, performance and theming, in development driven with an AI agent
- Period
- November 2026
- Team
- Intelligence Industrielle product team
- Stack
- Next.js 16, React 19, TypeScript, Tailwind CSS 4, Generative AI
- every field points to its page and sentence
- Cited
- priority comes from explainable rules, not the model
- Rules
- always validates before export
- Human
- PDFs, scans and photos, in both languages
- FR / EN
Try it
From a pile of reports to a batch of work orders.
Start the reading: reports get sorted. Open one, change a signal and watch the priority recompute; validate, then export. This is the principle with made-up rules and data, not the internal tool.
01The problem
At an industrial client, maintenance vendors send their service reports: PDFs, scans, photos, handwritten notes, in French or English. One person reads them one by one; the backlog grows and urgent recommendations get lost.
02What I built
- The review screen: reading the report, the proposed work orders and filing.
- A realistic demo mode, with edge cases and reading spread over time.
- A reading badge and notifications to follow reports in progress.
- Locating citations in the page image when the report is a scan.
- Performance, code cleanup and traceability; fixes and much smoother scrolling.
03Architecture
- 01DropEmails, .zip, folder
- 02AI readingCitations: page and sentence
- 03RulesExplainable priority
- 04Review and exportA human validates
04Decisions
- 1AI locates, rules compute: priority comes from explainable rules, not from the model.
- 2Guardrails: a stopped asset or a health and safety risk never comes out as low priority.
- 3Every field cites its source, page and sentence, boxed on the document.
- 4Québec’s Law 25: the AI notes that a technician was present, never their name.
- 5A human always validates before exporting to the maintenance software.
05Where it stands
In progress at Intelligence Industrielle (November 2026). Presented here as in an interview: no client or product name, no internal figures or data; the demo shows the principle with made-up rules and data. Development driven with an AI agent. Happy to go into more detail in an interview.