The Bag Alert

Architecture Diagram

What's the problem?

The moment an individual hits 18, they are suddenly faced with unfamiliar financial responsibilities. Those duties grow in urgency and complexity as these young adults take on loans, work their first jobs, and live their lives. Our solution: A platform that reduces the daunting details of fiscal responsibility into lessons that engage complete, long-term understanding.

Where did we start?

Once our team decided to teach financial literacy, we agreed that the most challenging aspect of any kind of learning was consistency. The easy fix: find something that you could and would do every day; an easy part of your routine that didn't feel like a drag. This led us to the American staple of starting your day: a coffee and a newspaper. An easy, relaxing way of learning about your world that millions of people could and would do without ever being reminded. The Bag Alert was born.

What makes it work?

Centering ourselves on accessible learning, the key features of The Bag Alert foster growth via rewarding interactions with the information delivered, and personalized difficulty to challenge the user without overwhelming them. The user is given five daily "articles" -- short, focused lesson modules tied to overarching financial topics (e.g. Savings, Investments, Credit). The modules' curriculum is powered by generative AI, creating a motivating, individualized learning environment. Atop generating the article content itself, the model attaches a multiple-choice question associated with the material, quickly cementing new knowledge with an incentive. Furthermore, they employ difficulty scaling-- offering beginners a high-level overview while generating more complex and detailed lessons for users who are better versed in a given topic. The user builds their streak by reading each of the day's articles, encouraging them to return tomorrow. Then, the user is presented with the Daily Game: a challenge scenario based on material from the day's articles. The user enters a free response solution to the problem imposed, offering direct creative involvement and reinforcing the article lessons. Afterward, AI evaluates the user's response, scoring and explaining the scenario. Each activity: answering the article question, responding to the daily game, and even bonus points from the streak, sum into XP points that add to the user's account level. Another simple reminder of the time the user has invested, encouraging them to continue growing and succeeding. Finally, to maximize accessibility, we incorporated multilingual articles and text-to-speech, so that The Bag Alert could deliver information to as many people as possible.

How did we do it?

Upon creating their account, users select the topics they want to learn. All account data including their topic selection is stored into AWS DynamoDB. They are later able to add more topics via the "Subscribe" feature. The lessons are structured topic-to-topic., meaning that each financial topic has its user level and multiple sub-topics. These are also stored in the user's DynamoDB entry. Of however many topics the user is "subscribed" to (topics that the user has indicated their interest in learning), the five lowest-level topics are used for the day's articles. As topic levels increase the more lessons the user takes with them, this system results in the user learning all of their subscribed topics at a relatively equal pace. To generate a tailored learning experience, we employed Claude 3.5 Sonnet v2 through Amazon Bedrock. We assigned a parameter to factor difficulty into article generation, based on the user's current progression. User's levels in a given topic were tracked by how many of its sub-topics they had completed thus far. Utilizing Amazon Translate and Amazon Polly lets us incorporate multilingual lessons as well as text-to-speech. The Bag Alert's infrastructure was hosted off of EC2, with its frontend and backend connected through Next.js and AWS Lambda.

Challenges and Solutions

Backend Challenges Coordinating the EC2 instance with DynamoDB presented challenges due to data formatting issues and logic inconsistencies. To resolve this, we manually formatted the data to ensure compatibility and integrated it into predefined routes, while rebuilding the logic for a seamless connection between EC2 and DynamoDB. Another struggle came about in finding the ideal prompt template to get consistent, accurate responses from the model. Our solution was using Claude itself for repeated prompt generation as it gradually optimized. Frontend Challenges Opening unique overlays for different articles posed a problem because of their different sizes. We handled this inconsistency by creating a uniform overlay regardless of the initial article cell size, and then adjusting the overlay based on class.