As a part of South-by-South-West 2019, Hypergiant and Vinesleuth partnered to create a kiosk experience that utilizes data from wine selections that span from all over the world to help the end user determine alternative wine choices based off of the taste pallete of the user. This pallete derives from an existing domain of taste qualities of which include the following: Sweetness, Acid, Spice, Fruitiness, Oak, and Tart. From this data, we are able to infer other wine choices from any region in the world that has a similar taste pallete.
The user flow of the application allows the user to explore wines by the Continent, Country, and Region. This drives the user to deep-dive into some of the favorite flavors all while passively learning and understanding how different regions around the world influence the components of wines. By using a generic map view, navigating to specific regions is designated by interactable zones that zoom the user into the subset of data and further filtering the wine selections. When the user identifies wine(s) from a region that they enjoy The Sommelier AI will infer a selection of wines at the global level that are similar in flavor profile of the selected choice.
A secondary feature of The Sommelier AI is purely based off of the users aesthetic preferences of wine labels. By utilizing a selection of various style label art, we tell the user a story of how shopping by label may not actually yield results that the user expects. As users shop with their eyes looking for labels that may influence what is expected of the flavor profile, such as bright bold colors, clean crisp line drawings, or purely typographical, not all wines return the flavors that are expected. When the user selects a label of preference, we bring up wines that share similar label properties but denote a drastic difference in flavor profiles. Some of which are completely inverse of the users selection.
As Cretive Technologist, my role's responsibilites started after the intitial design wireframe. In a short one week timeframe I was held accountable for building all aspects of the application as well as collecting, cleaning, and servicing a model to be used for the wine preference by label art [as mentioned above].
The collection and cleaning portion of the data set involved an external python tool written to identify and isolate key aspects of the supplied wine labels. This involved building a three tier software solution that scraped for content, cropped, and sorted images based on very broad conditions [Number of Colors, Line Weight, Histogram] with minor by-hand adjustments for classification. Each label had to be associated with the proper Index ID that is tied into VineSleuth's database of wine, where upon user selection all relative wine labels were parsed [predicted] and returned a list of IDs to query and load within the application.
The Unity application was built under the atomic design principles to handle for the fast turnaround time required. Being that it was an interface-heavy application, each component was built to be scalable across the board, with interactive elements being handled as class extension functions for easability of use in the code base. This enabled myself to have just the right amount of time to add in the necessary polish to handle for interface transitions, interactions, and optimizations.