AI Presentations
AI-assisted presentations: describe a topic, the AI proposes an outline, writes and lays out the slides; you adjust, then present.
- Role
- Main developer since September: features, fixes and tests, in development driven with Claude Code
- Period
- September – November 2026
- Team
- With the Intelligence Industrielle product team
- Stack
- Next.js, React 19, TypeScript, Tailwind CSS, Playwright, Generative AI
- first: the AI proposes, you reorder, then it writes
- Outline
- co-editing without losing a colleague’s work
- Live
- tracked translation of the whole deck
- FR ↔ EN
- or PowerPoint turned into editable slides
Try it
From topic to presentation, in a few clicks.
Propose an outline and reorder it, let the writing happen, then edit with AI, translate and delete several slides at once. This is the principle with made-up data, not the internal tool.
01The problem
A manager has to prepare a presentation “for this afternoon”, with no training and no designer. The tool turns a request, and a document if needed, into a presentation in the company’s style, which you then correct, translate, share and present.
02What I built
- Direct text editing on the slide.
- “Outline first” creation: the AI proposes an outline you reorder before writing.
- Live collaboration: presence, locks and comments.
- Converting a PDF or a PowerPoint into editable slides, and tracked French ↔ English translation.
- Cheaper AI edits: small replacements rather than a rewritten slide.
- Roles, trash, the interface redesign, the English version and iPhone fixes.
03Architecture
- 01RequestTopic, document if needed
- 02OutlineReorder before writing
- 03AI writingIn the background
- 04Adjust and presentAI edits, sharing, presenter mode
04Decisions
- 1Co-editing without losing anything: per-paragraph locks and three-way merge; “Undo” never erases a colleague’s work.
- 2For a small edit, the AI returns replacements applied all or nothing, instead of rewriting the slide.
- 3Realtime only signals; the database stays the single source of truth.
- 4Tiered automated tests, and mutation testing to check that the tests really catch errors.
05Challenges solved
- Very large decks: split uploads and batched export, so one stuck slide no longer stops the whole export.
- A PDF turned into editable slides: the text is placed back over the page, with OCR when the PDF is an image.
- Safari on iPhone ran out of memory: only one slide stays “live”, the others become images.
06Where it stands
In use internally at Intelligence Industrielle. Presented here as in an interview: no product name, no internal figures or data; the demo shows the principle with made-up data. Development driven with Claude Code. Happy to go into more detail in an interview.