ev.ai

Adding multiple characteristics
Available countries
User to describe target group
Choosing which market to evaluate
Loading Screen
Evaluation Page
Landing Page
Time to evaluate!

Inspiration

Big corporations like McDonald's and IKEA are able to succeed because of their able to invest deeply in target marketing campaigns and product localization in different countries. However, small e-commerce sellers are unable to do so with their tight profit margins and lack of economies of scale. We hope to address this issue by introducing a platform to support small businesses in expanding to new markets overseas, without exhausting financial resources in conducting real-life market research and user surveys.

In addition, data privacy laws can prevent businesses from being able to effectively use customer data in their market research. Our solution enables businesses to use synthetic customer profiles instead to test out their ideas without using customer data.

What it does

ev.ai is a platform which allows e-commerce sellers to bring their products to the world. Through our platform, sellers are able to take full advantage of LLM agents to efficiently gather customer feedback and actionable insights, which they can use to tailor their products and advertising to cater to a diverse range of audiences.

By generating synthetic user profiles, we are able to gather large amounts of data without collecting real-life user data, which reduces the privacy concerns associated with working with user data.

How we built it

Framework: Front-end:

Next.js, Chakra UI

Back-end:

FastAPI
Anthropic API

We used real life consumer data from an e-commerce Kaggle Dataset (https://www.kaggle.com/datasets/datascientist97/e-commerece-sales-data-2024) to enable the LLM to understand the e-commerce markets and the demographics of potential customers.

Based on these examples, via few-shot prompting, we allow the LLM to perform in-context learning and generate synthetic users with a variety of descriptions and characteristics, which mimic real-life demographics based on given "traits" (eg. country of origin, sporty, students, professionals, etc.). The LLM hence generates multiple user personalities, which act as individual LLM agents. These LLM agents are then surveyed on a selected product, and each agent gives customized feedback based on the agent's synthetic profile generated earlier, as well as multimodal information from the product (product title, product description, product image). The feedback includes both positive attributes of the product that stand out to the agent, as well as areas for improvement. On our website, users can view the individual feedback provided by a selected number of agents.

We then use all the feedback consolidated and aggregate the data, to provide a holistic view of the product, together with its positive and negative attributes. This is enabled by a website on the front-end to display the summary of the data.

Challenges we ran into

Lack of experience with long-term software engineering projects
First hackathon experience as a group
First time working with Multimodal Large Language Models
Obtaining reasonable training data

Accomplishments that we're proud of

Coming up with the entire idea for the project on our own, and implementing it in less than 24 hours
Experimented with multiple frameworks and methods and learnt a lot about software engineering and GenAI
Our team members complemented each other well