Wing AI

Inspiration

E-commerce, known for its convenience, has transformed the way we shop, yet it can feel isolating and lacking human interaction. However, the integration of generative AI and LLMs brings superhuman reasoning and conversational abilities to customer service, enabling interactions with highly intelligent representatives who personalize experiences based on shopper profiles. This also provide invaluable data to businesses, optimizing marketing and inventory for tailored product recommendations. By bridging the gap between convenience and human touch, AI-powered e-commerce offers an enhanced and personalized shopping experience, combining the best of both worlds for customers. Moreover, we - as fervent adopters of ChatGPT - founder ourselves using the site to ask questions like "what should I buy for a baby shower?" "how to build a deck?", etc, begging the question, why can't this conversation be hosted within the website itself where the user can have a far more enriching shopping experience, and the store can gain invaluable insights into their customer profiles, push out personalized marketing, and improve their inventory decisions based on customer sentiment.

What it does

Wing is a conversational search vendor for e-commerce platforms that transforms how shoppers interact with the internet and how businesses understand their customers. Wing enables customers to naturally converse with an AI-shopping assistant from ideation to checkout - able to help them through any step of the process and answer any question they have. Wing platform enables businesses to gain a deeper understanding of their customers’ thought processes to build a robust and ML-oriented customer profile, push individualized product recommendations, and gain inventory management actions. Wing offers an unparalleled conversational search experience integrated directly into your website, along with advanced product discovery & bundling, optimized inventory decisions based on analytics, and effective cross-selling and upselling capabilities. Our solution delivers value to customers and storefronts alike, setting us apart from the competition.

How we built it

We developed Wing with a Node.JS application as well as a Flask backend. We receive data in CSV or similar formats, embed the data into a Pinecone vector database, and then perform neural search on the embeddings when a customer queries. We also have a live deployment that is embedded within the HTML/CSS of a demo website. One of our biggest advantages over any possible threat is that we have Reinforcement Learning through Human Feedback, this means users can rate how they liked their conversation as well as Wing logging which products they click on and pursue to purchase. With this information, Wing trains a Customer profile module that is a way for the business and our LLM to understand who our customer is in a very rich, detailed, and concise format - leading to better recommendations. Finally, we have an analytics engine that uses GPT to understand customer profiles and "boost" products into the profiles of customer's who are likely to interact with them based on past product history and confidence of the model that they will interact. Finally, Wing allows businesses to customize and personalize emails and promotions to customers, where customers will only receive promotions for products that they are likely to interact with and stated in a manner that directly appeals to who they are, what they are looking for, etc.

Challenges we ran into

Originally, we had a very structured MySQL database upon which we trained a MindsDB Hugging Face Table Querying Model on in order to derive insights into whether or not a customer or group of customers would like a certain product or not. Unfortunately, this did not work out after repetitive attempts and we pursued a nearest k-neighbors approach. Another challenge we ran into was related to our focus on improving the customer experience on the site. Our approach to this was directly linking images and appealing hyperlinks on the chatbots response if they are products the store offers. This was difficult due to streaming the data as it goes through a bi-layered LLM infrastructure while maintaining integrity of the data to eventually provide links to the user.

Accomplishments that we're proud of

We are extremely proud of being able to develop a lightweight infrastructure that allows significant benefits for both the customer visiting the website through a revolutionary way to interact with the business' products, as well as a novel approach for businesses to extract natural language analytics, develop personalized marketing and product suggestions, as well as inventory management to maximize revenue.

What we learned

Through the development process, we learned the importance of making a product that is scalable and defensible in the long run. It is not difficult to develop a GPT wrapper that simply responds to customer questions, but we realized Wing's strength and unique value in being able to develop a comprehensive, rich, and efficient customer profile that speaks to their buying behaviors, preferences, etc, meaning they have a truly personalized experience on the website and the store has more efficient touchpoints with them. We also realized the importance of understanding that limiting our development is important for adoption. Although there are numerous verticals for which genAI and LLMs can be deployed, starting with a specific solution that works for both businesses and customers and can be easily implemented, learned, and utilized is crucial.

What's next for Wing

Currently we maintain an RLHF algorithm for increased customer personalization; however we acknowledge that the future bears a concern regarding individual privacy. Therefore, we aim to address the issue by introducing a distributed network of clients that share access to a customer's data across platforms while enabling the customer to maintain complete control and oversight into which stores have access to their data. In this, the stores will be able to compound their revenues through heightened personalized recommendations by building a rich customer profile, and customers will have complete knowledge and control of the whereabouts of their data, addressing the data privacy concern.