Language, image, and voice models are the engine, and the product.
Decklyn could not exist a few years ago. Writing structured slide copy, designing a matching visual for each slide, and narrating it in a natural voice, together, is only possible with modern AI. That is not a feature. It is the whole product.
A language model writes the deck
You give a topic and audience. A language model returns a full deck as strict, validated JSON: a title, a subtitle, and each slide's title, bullets, spoken narration, and a short visual description. We validate every field with a schema, so a malformed response degrades to a safe fallback instead of rendering garbage.
- Structured output, validated with zod
- One draft, not a chat, so it is fast
- Model-agnostic: we can swap the model without touching the product
An image model designs each slide
For every slide, the visual description the writer produced becomes a prompt for an image model, wrapped in one consistent house style: bright, high-key, on a white background, on your brand color, no text or people. The deck looks designed by a single hand. Quality is pinned per plan, never left on auto.
- A generated visual for every slide
- One house style across the whole deck
- Regenerate any single slide you do not love
A voice model narrates it
Each slide's narration script is sent to a voice model, which returns natural, professional speech in the voice you chose. That is what turns a deck into a talk: you can present it live, slides advancing with the voice, or share it as a narrated deck someone watches on their own time.
- A real, professional voiceover
- Slides advance in time with the voice
- Falls back to on-screen notes if voice is off
How we build with AI
Draft, do not deceive
Decklyn writes from your topic and marks its output as a draft to review. It does not invent statistics or citations and present them as real facts.
Model-agnostic
The language layer is a thin, swappable interface. As models improve, Decklyn moves with them, no rewrite required.
Keys stay server-side
Every model call runs through our server routes. Your keys, prompts, and content never touch the browser.