Change Your Financial Lifestyle with BudgetBuddy

BudgetBuddy

Inspiration

BudgetBuddy was born out of the need for a practical, immediate solution to manage and optimize personal finances. Everyone faces financial concerns in life, but finding a satisfying service can be challenging. However, the idea of using generative AI sparked our belief in creating a fundamentally different financial management service.

Therefore, we defined the customer problems we wanted to solve using AI as follows:

Everyone desires economic growth or stability – Common & General Needs.
Individuals have diverse economic situations – Private & Individual Situations.
Traditional financial services only offer generic advice and fail to deliver immediate effects.
Financial knowledge is often difficult for average users to understand.

We identified the following tools from a technological and planning perspective that could solve problems:

Using LLM allows for personalized analysis with less time and cost.
Financial management focused on cost reduction can bring immediate effects.
Continuous and practical improvement in spending habits can be induced through a persona created by LLM.
LLM enables the provision of financial outputs that are easy to understand, tailored to the user's level, through natural conversations.

Utilizing these market insights and technological tools, we developed BudgetBuddy, a real-time, highly personalized financial service.

What it does

BudgetBuddy is a highly personalized financial management service that leverages a Language Learning Model (LLM), specifically GPT-4, to deeply understand and respond to users' financial situations. Its features include:

Personal Financial Assessment

Analyzes spending patterns and financial status, integrating dispersed financial data into personalized reports.
Delivers analysis results in a financial language that is easy for users to understand. Spending Behavior Control
Provides real-time interactive feedback on potential unnecessary expenditures, with intervention to guide users towards savings.
Utilizes a friendly persona for effective communication. Continuous Personalized Financial Management
Continuously monitors financial status and adapts advice accordingly.
Has the potential to expand beyond expense management to include guidance on increasing income.

How we built it

Data Ingestion

Integrated a pipeline for ingesting financial data from various institutions into Databricks via APIs.
Enabled real-time tracking of user spending and analysis using an LLM, with data streamed into Databricks.

User Interface (UI)

Developed the UI using Streamlit for user-friendly interactions with a virtual assistant, named Charlie.

Data Analysis

Utilized the GPT-4 Turbo model from OpenAI for analyzing spending patterns and expenditures.
Incorporated the Code Interpreter feature of GPT-4 Turbo for in-model data analysis and user feedback.
Transferred financial data from Databricks to GPT-4 Turbo for LLM-based response generation.

Implementation

Employed OpenAI's Assistant API for seamless integration of user inputs and LLM responses.

Prompt and Persona Design

Created a virtual persona, Charlie, as part of the LLM Instruction to interact effectively with users.
Write a detailed prompt describing Charlie's personality and characteristics.
Designed functional specifications as prompts for accurate service responses.

Challenges we ran into

During the development of BudgetBuddy, we pondered solutions to two major issues as follows: The first challenge was about the legal regulations and methods for collecting actual financial data, as they vary by country. This posed the question of how to incorporate these varying elements into our service. In South Korea, the government provides an integrated financial data service called Mydata. Therefore, we assumed that data would be saved in Databricks in a streaming format via an API from the mydata service. It was not easy to analyze users' financial data using only the LLM itself. The responses provided by open-source LLMs or existing LLM models made it difficult to understand the actual process of data analysis. However, with the introduction of OpenAI's GPT-4 Turbo model in November, which includes a Code Interpreter feature, it became possible to implement analysis using the data provided to the LLM model. Moreover, we integrated this feature into the actual service functionality, allowing the Assistant API to provide users with responses based on the analysis performed through the Code Interpreter.

Accomplishments that we're proud of

We are particularly proud of:

1. Feasibility Confirmation as a real service: VVerified the real potential of the service by deriving useful analysis from actual spending data, obtaining immediately actionable results.

2. Charlie's Effective Communication: Successfully developed Charlie as a persona that goes beyond a simple chatbot, embodying a friendly and kind personality. This achievement led to the realization of natural and understandable language in the character's tone and manner, fully aligning with our objectives and enhancing user engagement.

3. Interactive Financial Service: Developed beyond one-way financial advice, effectively improving users' behavioral habits through an interactive feedback function.

What we learned

As we planned, designed, and developed this service, we utilized various generative AI services. The help of these generative AIs brought about a level of productivity that was unimaginable before. Our service is not just a simple text input/output LLM service; the recent GPT-4 Turbo model has provided us with an opportunity to gauge how much more creative and versatile a multi-modal service can be. We believe that generative AI will make it easier to implement hyper-personalized services based on individual data.

What's next for The important thing: "Unbreakable Heart"

Looking ahead, BudgetBuddy aims to:

1. Integrating the service, currently developed as an MVP, into a fully functional product ready for actual delivery.

2. Go beyond reducing expenses to expand features that offer suggestions for increasing income.

3. Refine our data engineering architecture to more effectively integrate dispersed personal financial data.

4. Diversifying the Charlie persona, for example, by purchasing the IP of renowned investors or financial managers like Warren Buffett to create personas capable of providing expert advice, and offering these to users.

Our team is dedicated to responding to economic challenges and nurturing a healthy financial lifestyle by creating a service that provides tangible and effective financial management for users."