This past year the Gretel team has been hard at work building tools to help engineers and data scientists build with safe data. This task is no small feat, and we quickly discovered that with all the hard work we put in, we longingly looked forward to our Friday afternoon happy hours. Through a series of meetings, team off-sites, R&D, and a lot of self-discovery, we finally built something that would allow us to work hard and play hard. Introducing the Gretel Bartender.

The epiphany

Hindsight is 20/20, and what seems so obvious now was entirely lost to us at the beginning. One night while sipping on cosmopolitans at the Casa Jondal in Ibiza (OK it was a Chili’s in Salt Lake City), our CEO – Alex Watson – looked disappointingly at his empty martini glass and blazoned, “That was delicious, but ’m getting tired of drinking the same thng all th’time. When’s the last time anyone came up with a newdrink?”. That question (though very slurred) sparked an idea in our minds. What if we could use our own tools to design an AI that would save us from the endeavor of having to drink the same boring drinks over and over again? We have the technology, so what’s stopping us?

This is the greatest thing since the invention of the keg stand!
– Florida State University student, Chad

The Plan

With our minds buzzing with anticipation and excitement, we set off to start building a solution to our very real, very important problem. We wanted to train a machine learning model on a large database of cocktail recipes and teach it to craft brand new ones. These new, synthetically generated recipes would be a perfect conglomeration of all the best drinks in the world, from the humble old-fashioned to the bold and daring Double Chocolate Martini. It would select all the best ingredients, ratios, and mixing techniques and unite them into wonderfully new and exuberant creations that most mixologists could only dream of creating.

I can’t believe you’re wasting money on this.
– Gretel’s Investors

The Challenge

Soon after its inception, problems began to emerge. The AI wasn’t performing very well during its training. As any boozy connoisseur will know, you can’t just mix a cocktail and send it off. You have to try it for yourself. After each training epoch, our model became tipsier and tipsier. As each iteration processed, it became more imprecise. It started to produce garbled nonsense. It wouldn’t stop talking about politics and putting in Jukebox requests for “Piano Man” at all the local bars. It wanted to plan a group trip with all of us that “We’ll totally go on this time, and I’m not just saying that because I’ve had a few. I mean it”. Eventually, we had to cut it off. We told it that there is important work to be done and that it was time to buck up; we’re so close. After our heart-to-heart chat, we finally began to see some real results, and boy, were they wonderful.

Just one more, it’s not like I’m driving tonight.
– Gretel’s ML Model after four gin and tonics

The Results

The results it served were better than we ever could have imagined. The model could produce an infinite number of brand-new cocktail recipes. All we had to do was suggest a base liquor, and it would fill in the rest. Feeling something with Vodka? How about a rootbeer vodka with grenadine in a highball glass with a salted rim (yummy)? Or an Irish coffee with champagne, red bull, and a lemon wedge (that will wake you up)? We were in wonder at the possibilities and knew immediately we had to make this available to everyone. So, in the spirit of giving and a desire to make the world a better place, we are opening up our AI to the world.

To generate a new cocktail recipe, all you have to do is tweet at our Twitter bot with your favorite liquor, and we will send you back a delicious recipe.

Closing

We hope you enjoy your April 1st with a synthetic cocktail or two. Here’s the Cocktail Ingredients dataset from Kaggle we used to generate our model. If you actually end up making one of these drinks, tweet it at us, and we’ll ship you something fun!

Cheers!