Client work
Pour AI: a sommelier with no attitude
Client work, designed and built by Riv Lux Media.
The barrier to wine is embarrassment. Information was never the problem.
People don't ask because they're afraid the answer will come with a lesson attached, or with vocabulary they're supposed to have already, so they buy the second-cheapest bottle and stop thinking about it.
Pour AI's whole positioning is the fix: your personal sommelier, no attitude required.
Two ways in, and the second one matters more
Ask a question. The obvious path.
Or photograph your meal. That's the one that changes who will use it. To describe a dish accurately enough to get good advice, you already need food vocabulary, and that's the same barrier again, one step earlier. A photo requires nothing. You point the camera and get an answer.
On a spec sheet, multimodal input would be one more line. Here it's the difference between a tool for people who already know how to ask and a tool for everyone else.
Tone is the product
"No attitude required" is a constraint. It goes on the landing page, and it also has to be enforced in the output. An assistant that answers a simple question with a paragraph about terroir has failed, however accurate the paragraph is.
That puts the design work in what the model is not allowed to do: lecture, assume prior knowledge, hedge into uselessness, or perform expertise. Warmth and brevity are functional requirements in this category, and they're harder to hold than correctness.
Scope is what makes it trustworthy
An assistant that will answer anything will eventually answer something wrong, confidently and in public. In wine that's worse than useless. A confidently wrong recommendation costs someone money, and their evening.
So the interesting engineering is the boundary: what the assistant answers, what it declines, and what it does when it isn't sure. We build AI features in this order. First we decide what "correct" means and how to measure it. Then we constrain the scope to something the assistant can be right about, and ship the fallback for when it isn't. Making it feel good comes last.
What this demonstrates
Multimodal input used to remove a real barrier. Tone treated as a product requirement and enforced. And an AI assistant scoped tightly enough to stay trustworthy in public.
Last updated