Multi-Agent financial coach and assistant

Starting from the very beginning, our team at Vstorm had to propose the Agentic AI architecture that supported the bold goal of building a universal financial advisor.

The agentic architecture

To support this goal, we decided to use the supervisor architecture, where the leading agent coordinates the work of the rest, splitting user’s request into separate tasks and giving them to more narrow, specialized agents. It can be compared to construction works, where the leading engineer is asked to build a house and knows the procedures who then calls builders, then electricians, plumbers, or other specialists to deliver the best completed work possible.

With the supervisor agent overseeing work on the prompt, the agents may get or not get a task to perform, launching either a chain of actions or one particular action which needs to be done by each individual agent, without the need to send notifications or use tokens by the those that remain unneeded in each particular instance.

In this case the team of specialized agents included:

• Supportive agent – an agent that is set to always find positive aspects and bright sides to a user’s financial behavior. The agent has access to aggregated data about the financial habits and patterns of people of various demographics, so it can compare the user’s decisions to most common ones. Due to this hard-wired enthusiasm, it cherry picks the best spending habits and chooses others to show when the user is on the right track.
• Cautionary agent – the direct opposite of the agent described above, this one is always worried and cherry-picks failures or red flags in the user’s financial decisions. The same as the above, this agent has access to data about the financial habits and decisions of particular demographics, so it can compare, analyze and spot what is wrong or needs adjustment.
• Assistant agent – Yeeld provides a handful of features including savings and investment tools. The assistant agent helps the user manage the app and account, powered with documentation that is accessed via specially tailored RAG, so the system will not hallucinate or confuse answers.
• Financial agent – this particular agent can read data from all external tools. The agent operates and can access the banking data and return answers based on them. It is also capable of generating graphics and visualizations, so the user may get one’s data in the form of a chart or graph.
• Action agent – this agent maps the user’s request to the app’s capabilities and returns possible actions to take. If the app is unable to perform a particular action, for example to provide a tool to buy shares of a particular stock, the agent also informs the user about the limitation.
• Conversational agent – this agent is used when the user is just running a casual conversation without requests which need to be processed by the rest of the agents. All the casual “hello,” “goodbye” or “thank you” messages are handled by this agent. To save on tokens, the agent is instructed to be polite, yet to respond when and that a topic is not of its competence to discuss.

The agentic ecosystem was built using LangChain and LangGraph technologies.

Gamifying spending

To further enhance the experience, our team at Vstorm delivered two games that support users in building better spending and financial management habits.

Swipe the bill

This game uses a Tinder-comparable interface of swiping left and right, but instead of potential matches, there are bills and spending information drawn from the user’s account. With the swiping motion, users may either approve or condemn the spending, reflecting on one’s financial decisions and looking for possible savings. After the game has ended, the user gets a report on how many payments were worth it and how many were not.

Are you smarter than the rest

Quizzer, the second game, lets the user contrast their spending habits with that of the comparable population. The user is asked whether one thinks that one spends more or less than other people on a particular good or service (food, clothing, electronics, services, etc.) and can compare their habits and expectations with the generalized data.

The effect

Our work at Vstorm was in delivering a fully functional Agentic AI workflow that orchestrates the work of multiple agents on user’s prompts to provide better financial advice and instructions on how to improve one’s overall financial health.

The app also uses non-standard notifications that can provide insights on the user’s ongoing behavior.

.