Multilingual, multimodal, designing an assistant that understands seven languages

Season one proved the appetite

BourbonIT drew twenty eight thousand recipe submissions and crowned forty five winners in its first season, a clear signal that people wanted to engage with the brand creatively rather than passively. The question for season two was not whether the appetite existed, it was how to reach a country that speaks dozens of languages and, for a large share of its audience, does not particularly want to type a recipe out on a phone keyboard.

The brief for season two

Britannia wanted more depth and genuinely more access, participation that felt effortless whether someone spoke Hindi, Tamil or English, whether they preferred to talk rather than write, or whether they simply wanted to share a photo of a dessert they already loved rather than describe it from scratch.

What multimodal actually meant here

We built BourbonIT 2.0 on Google Cloud so a recipe could start from any of three entry points, speech, a photo, or a shared link, and the platform would generate a personalised Bourbon recipe from whichever one someone chose. It understands seven languages and counting, and a real time AI cooking assistant answers questions mid recipe, from ingredient substitutes to exact mixing times, the kind of question a home cook would normally ask a more experienced relative rather than a search engine.

Google Gemini does the actual work underneath all three paths, vision to recognise a dessert from a photo, speech recognition across multiple accents for voice input, and language generation to turn any of those inputs into a coherent, personalised recipe. Every entry becomes an AI enhanced, branded recipe card built to be shared rather than filed away, and community voting powers a People’s Choice award that gives the platform a reason to keep coming back to after the recipe itself is finished.

What we learned about language and trust

The clearest result from testing this with real users was not a feature preference, it was an emotional one. Letting someone speak in their own language, in their own accent, changed how much they trusted the platform with something as personal as a family recipe, far more than any single interface improvement did. Recipes are inherited, not just cooked, and an assistant that could not be spoken to naturally was always going to feel like a translation of the experience rather than the experience itself.

That is the detail we now carry into every multilingual product we build, voice input is not a convenience feature added onto a text first product, it is frequently the entire reason someone trusts the product enough to use it for something personal.