Inspiration

In the early days of my career, I took on a software role at an insurance agency. My responsibilities spanned data cleaning, software development, and occasionally standing in for an insurance agent during unforeseen absences. This experience provided me with an intimate understanding of the insurance industry and its accompanying challenges. One vivid memory involves a long list of phone numbers I had to dial to resolve claims. Each call could stretch beyond two hours, particularly when handling ubiquitous claims. The monotony of waiting on hold, often to the tunes of Anthem Blue Shield Blue Cross, remains etched in my memory. I felt the inefficiency acutely, thinking of all the tasks that awaited me.

What it does

Data Wall offers a sophisticated chat interface, guiding users through the intricacies of their insurance and related claims. Utilizing a conversational Natural Language Processing model, fine-tuned with comprehensive insurance data, our platform engages users in a detailed conversation. The model intelligently inquires until it has garnered sufficient information about the user's specific concern.

Once equipped with the necessary details, at the user's command to "make a call," the model takes the initiative. It autonomously contacts the relevant insurance company, representing and advocating for the user's concern. Upon conclusion of the call, our system promptly sends a text message and email to the user, apprising them of the outcome and delineating any subsequent steps that might be needed.

How we built it

To automate these tedious processes, we leaned heavily into Natural Language Processing and Speech-to-Text technologies.

• Backend: Crafted using Python and Flask.
• Frontend: Developed in React and Typescript.
• Chat Interface: Powered by Google's FLAN-UL2 model.
• Speech-to-Text Conversion: Leveraged OpenAI's Whisper model.
• Hosting & Training: Amazon's EC2 instances were instrumental.
• Call Automation: Made possible through Amazon Connect.
• Data Management: Streamlined with Amazon Kinesis. In prioritizing our user's privacy and aiming for HIPAA compliance, we hosted our models on-premise, rather than relying on third-party solutions like ChatGPT. Our current setup provides us with 40GB of GPU memory. While this isn't the pinnacle of scalability, it aligns with our dedication to user privacy and data protection.

Challenges we ran into

Surprisingly, our main hurdles weren't rooted in the complexities of training our models or delving into machine learning intricacies—thanks to our team's solid background with Large Language Models and prior machine learning engagements.

Our primary struggle revolved around comprehending the nuances of the data pipeline, specifically the flow from Amazon Connect (our telephony service) to our Speech-to-Text model, Whisper. Visualizing the movement of data through Amazon Kinesis posed another challenge. Without a transparent view of the data as it journeyed through various stages, we found ourselves at a disadvantage. Being privy only to the final result can be limiting. A lack of insight into intermediate steps often left us questioning the integrity and quality of our data.

Accomplishments that we're proud of

Our accomplishments span both technical prowess and personal growth. While our team's forte has traditionally been back-end development, we ventured into the realm of front-end by mastering React—a journey that was as enlightening as it was challenging. We realized that front-end development isn't just about pure logic; it often requires a balance of aesthetics, user experience, and subjective decisions that differ vastly from the "cold logic" of back-end coding.

Equally commendable is our in-depth exploration of AWS services. Navigating AWS's intricate web of services felt like solving a complex puzzle. We're especially proud of demystifying the mechanics of AWS's data pipelines and gaining insights into real-time data streaming from live calls to our platform. This journey through AWS was a testament to our team's adaptability and eagerness to embrace new technical challenges.

What we learned

Our journey illuminated the pivotal role of infrastructure in the success of a project. We realized that managing infrastructure is far from trivial, and its importance cannot be underestimated. Designing robust, scalable infrastructure is both an art and a science. While the temptation often lies in immediately diving into coding, we recognized the perils of makeshift, "spaghetti" structures. Conversely, we discovered the immense rewards of thoughtful planning and design. This lesson underscored the value of preparation and strategy over impulsive execution, an insight we'll carry forward in all future endeavors.

What's next for Data Wall

Short Response:

We're opening our digital doors! It's time for users to experience the revolution that Data Wall offers.

Long Response:

As we eagerly anticipate welcoming more users, our immediate focus lies in expanding our capacity. Our current constraint is our EC2 instance, and we're diligently exploring innovative solutions to scale operations. The goal? Handle an influx of calls and amplify time savings for our users. While we gear up behind the scenes, our landing page acts as the gateway for potential users. Here, they can sign up and join an ever-growing queue of individuals eager to harness the power of Data Wall.

Built With

• amazon-connect
• amazon-ec2
• amazon-kinesis
• flan-t5
• gpu
• javascript
• python
• pytorch
• react
• tensorflow
• typescript