T-Care

Inspiration

In general, it is time consuming to navigate through the web because the web doesn't know what the consumer wants. It leads to many people wasting time creating tickets that don’t address the right problems and making complaints that fall on deaf ears, or they go in person to the store instead of getting the help they need online. Spending hours and hours on customer support calls is one of the most disliked ways of spending time, and often leads to unhappy customers, which impacts businesses. For enterprises, investing time and resources into improving customer service benefits not just the user. It also allows for a more efficient workflow of assistance, which allows more customers to get their needs fulfilled, improving the business’s success.

What it does

T-Care is a customer service tool for customers and employees that helps consumers navigate effortlessly to get the right products and services they need through interactively talking with a customer support AI and an active and precise feedback loop using CV recognition (Nemotron) for sentiment analysis, which provides live agentic updates on the page for personalized support. This data is given back to the employers on their side and compared to competitor data online using a Web Agent (Parallels & Gemini) which generates summaries and graphs of what competing enterprises look like, and where the client succeeds or falls short. There is also network infrastructure for the client to check when servers go down, are impacted, and what support the users need. The users get live updates through text message SMS on their network status if there is an outage.

How we built it

We used React.js for the front end, with Tailwind for the UI. On the backend for auth and network data storage, we used Supabase. For the CV sentiment analysis and central AI configuration, we used Nemotron. We used VAPI for interactive and responsive voice agents, Gemini & Parallels Web AI Agent for the Admin dashboard to scrape the web and view and compare feedback from users compared to competitors. We used Nodemailer for live SMS updates to users. For the live network coverage map with outages and problems reported, we used Mapbox.

Challenges we ran into

With the use of multiple API calls to AI agents, oftentimes models would hallucinate and incorrectly display or update the page in accordance with the user’s requests. This was solved through clever prompting of our agents, which allowed for the safest outputs with as many guard rails as possible to eliminate hallucinations. Additionally, creating a CV model that could analyze a user’s emotion and sentiment in real time and update accordingly to change the frontend and display what the agent predicts the user would like to see was extremely challenging, but we tackled it with robust weighting and learning from other examples of open source recognition models.

Accomplishments that we're proud of

Building a robust, interactive, and personalized agent that can be used as a service for any user regardless of their knowledge of how to navigate the web or access customer service, cutting time and increasing efficiency is something we are very proud to be able to build.

What we learned

We vastly improved at skills that we are not really familiar with, such as extensive CV modeling and mapping to sentiment, as well as integration with multiple APIs and AI implementations, to create a predictive and reactionary environment.

What's next for T-Care

We would love to pitch our agentic customer service idea to small businesses across Texas and Georgia (where our team is from). Using this product can potentially help those with lots of online traffic and e-commerce make their product better.