Wondering what wine to pair with Hainanese Chicken Rice, a Blue Cheese Burger, Baba Ganoush, or an Ice Cream Sundae? Pioneering food and wine pairing website, try.vi could help. Years ago, a glass of cheap and sweet rosé on a hot summer’s day made John Kaminsky fall in love with wine, a passion that eventually inspired the creation of this innovative pairing tool – a labor of love launched in 2023. These days, the CMS-certified sommelier and software developer’s rosé tastes run to Bandol and Tavel – and to sharing deep-dive, global sommelier-driven wine and food pairings that make that expertise accessible to a broad audience.

In the first of our food and wine pairing series, Westgarth Wines spoke to John about try.vi, how the tool generates its diverse and tasty pairing suggestions, and discovered how it lives up to its slogan: “A world of wines to try”. Plus, for food and wine lovers who are also techies, there’s a bonus section focusing on more technical aspects of try.vi.

First things first. There are thousands of food and wine pairings on Vi. What’s the last one you looked up for yourself, and what made the pairing work?

Kimchi and Alsatian Gewürztraminer. One of the fantastic opportunities with Vi is to work with sommeliers and wine professionals from all over the world, and a number of the somms are actively working in Asia. For this pairing, there is something about the spice and fermented character of the Kimchi that contrasts spectacularly with the floral, tropical aromas and the texture of the Gewurz.

The Try.vi homepage

Can you summarize for our readers what Vi does for wine lovers?

It collects pairing recommendations from experts and makes them accessible to non-experts who love exploring food and wine. The single most unique feature of Vi is that it brings you to interesting, underrated, and unusual wines and pairing suggestions quickly – and we explicitly include at least one “unusual or unexpected” wine in each set of recommendations. Food and wine pairing is a notoriously challenging and subjective topic. There have been a variety of attempts to quantify what makes certain pairings work by deep analysis of food and wine components, down to the molecular level. This approach simply doesn’t work – the aromatic and flavor compounds in wine are too varied and the interactions too complex. Wine and modern cuisine are human inventions, so in my mind, it makes sense to specifically capture the human experience of food and wine pairings.  

We’ve just inputted “Pan Seared Salmon” into Vi. The top suggestion we got was Willamette Valley Pinot Noir, along with nine more. What went into this recommendation?

Generally speaking, the first ten wines you see on try.vi for any dish will be a literal presentation of the sommelier team’s recommendations for that dish. By gathering recommendations from dozens of somms from different regions, we get a diverse initial set of options. We also have some logic that ensures some diversity in the recommendations. For example, Pinot Noir is a popular pairing with most salmon dishes, and we have many regional variants of Pinot Noir in our list of wines. So, the site will suggest a couple of the most-recommended Pinot Noir options, but then move on to other wines.

In a few words, what is Vi most definitely not?

I dislike when people think of it as an “AI sommelier”, with the implication that the recommendations are synthesized from masses of unvetted content, or that an algorithm is making the decisions. Vi is a tightly curated repository of professional wine expertise.

Speaking of sommeliers, sets, and aromatic/ flavor compounds et al, what kinds of experts are on the Vi team?

It’s a particularly accomplished group that has contributed to this project to date. The core technical team includes two PhDs, and the wine team includes sommeliers with experience at 3-star Michelin restaurants, Master of Wine candidates, as well as experts focused on specific wine regions. The three main areas of expertise are food science, wine, and technical.

Wine pairing suggestions for one of hundreds of international cuisine dishes on Try.vi.

As well as suggesting expert food pairings, the tool provides alternatives to help wine lovers explore new grapes and styles. We just searched for Blaufränkisch and were presented with 10 other wine options alongside the Austrian wine. How does Vi do this?

The simplest way to think about these alternative wine recommendations might be “if I were blind tasting, what other wines would I likely confuse this with?”. But, we also try to widen the aperture a bit further than that, to find wines that have some attribute that would make them appealing for fans of the original wine. The underlying approach to categorizing wines is loosely based on the same kind of tasting grid used by WSET or CMS students, but with a variety of additional dimensions that add depth and capture more traditionally qualitative nuance.

Following on from this, what’s the last wine alternative you looked up for yourself on Try.vi? Did you enjoy the choice/s offered?

That previously mentioned Gewürztraminer pairing sent me down an Alsace rabbit hole and brought me to Edelzwicker. Combining the aromatics of Gewurz and/or Muscat with the additional backbone from Riesling and/or Pinot Gris makes for a tremendous combination, especially for the price.

The site is super user-friendly. How did the team manage this?

A lot of thought went into it. I’d say that the website interface is designed to get you useful recommendations as quickly and intuitively as possible. Most other websites that have food and wine pairing recommendations require you to read quite a bit to understand and absorb a recommendation. We tried to create something more functional and intuitive, with multiple pathways to address different needs.

What kind of feedback have you had from users?

We hear that it gives people confidence to try something new, and they are often delighted by how their horizons are expanded when they do. I heard recently that the staff at one of the largest wine shops in New York City regularly uses the site to help find unique pairing recommendations for customers.

Have you had any surprises around usage?

One unique benefit of being able to see “behind the curtain” about how people use the try.vi site is seeing how traffic to various dishes ebbs and flows as it relates to seasons, holidays, trends, etc. On the subject of chicken, I had never heard of Coronation Chicken until King Charles was coronated in 2023, or “Marry Me Chicken”, which is TikTok famous and remains one of the most popular dishes on the site.

Finally, for our foodies, what’s your favorite food and wine pairing suggestion?

Would it be hurting the cause of Vi if people know that Champagne pairs with everything and that it could be the only wine you’d really ever need?

John Kaminsky

Thanks to John for this dive into try.vi. Why not tell us about the pairings you discover on Vi on our socials?

If you’d like to learn more about the technological landscape around the project, grab a glass of wine and read on.

While Vi isn’t an “AI sommelier”, does this technology play a role?

There are AI components to the project, but we try to keep them on a very short leash, in order to ensure the recommendations are faithful to the intent of the sommelier team. There are four main places where AI is in the mix:

Less common recommendations: When you begin using the tools to customize your recommendation, you may go beyond the boundaries of our sommelier recommendations. For example, if you say you’re specifically looking for an oaked Italian red under $50 to pair with a certain dish, we may only have a few direct recommendations, but the AI model can use our wine similarity data to come up with more recommendations where there is a high confidence in the results. So, for that Italian red, if Brunello di Montalcino is a top recommendation but outside your price range, it may steer you to a Vino Nobile di Montepulciano or Chianti Classico. Or if you step your target price range down further, it’ll lead to one of the value reds of Portugal. During development, we went through a number of cycles where the sommelier team carefully checked the soundness of these extended recommendations.

Text explanations: In the initial version of the site, we worked with a wine and food writer to write up descriptions for why each of the pairings worked. It took four months of full-time writing to write out the explanations for about 500 dishes. As the site has grown (it now features several thousand dishes), we’ve switched to using a large language model which is fed very detailed context on the pairings, and then it writes out the reasons why the pairings work in paragraph form. Candidly, I don’t like this because I think it gives the overall site an AI veneer, when what we’re really trying to do is showcase unique human recommendations. However, writing out a couple hundred thousand paragraphs of description text just proved to be too much work to approach manually, at least for the size of our team.

Incorporating more details about dishes: We’ve augmented our original food data set with additional context from the latest large language models. So, for example, we can augment the core flavor attributes with important additional historical and cultural context for each dish, including synonyms, cultural variations, and related dishes etc. This is something that we did manually for the first version of the site, but we realized that completing this process for thousands of dishes would be economically prohibitive, and AI makes it possible to automate effectively.

Search: Finally, we use some AI components to interpret search queries to find the closest relevant match when we don’t have the exact dish someone is looking for. What’s interesting is that our AI model has the ability to interpret and synthesize totally custom recommendations, for example, when various ingredients are combined in different ways. We have not turned that feature on for site visitors, though. The reason is that if you were to make weird requests (pairing for chicken and maitake mushroom ice cream?), the model will find the wines that seem to go best with those flavors. But, the answer may not make sense in the real world, and that would reduce the credibility of the site’s results. This is a broader problem with AI that is still being worked out - it doesn’t always know when an answer is reasonable or not. It's why there was one point where if you asked Google how many rocks you should eat in a day (which you can do), it would respond about the mineral intake benefits of eating one rock per day.  

How does the Vi team work?

Vi’s food science team develops a data structure for creating a reasonable framework with which to understand the composition and flavor profiles of dishes. Our wine professionals develop a framework for describing typical wine styles and create pairing recommendations, and the technical team members capture this expertise and transform it into data, along with developing a useful interface from which to explore it.

What are your future plans for Vi’s development?

We’ve built an API that will allow retailers or other businesses to integrate recommendations into their site directly, and are working through the best way to roll this out. There’s also interest in offering native iOS and Android apps. Meanwhile, we’re just delighted that people who may not be pairing experts are getting value from the site!

Want to read more? Take a look at some of our other blogs:

• • First wine memories: Bordeaux winemakers open up
  • Malbec World Day 2025: Food pairing ideas from Catena Zapata, Viña Cobos, and Achaval Ferrer
  • Stephanie Wangui: Swapping the corkscrew for vineyard life in Burgundy