aurum.ai

Inspiration

Understanding how my money moves has always been frustrating for me. Existing apps have poor classification, no granularity, and lack smart reconciliation. They cannot track my expenses with friends, my habits, and break apart large purchases into smaller chunks. e.g, If my friend pays for dinner and I have to venmo him 30$, that is not tracked! Or, if I make a 100$ purchase on amazon that includes a hair dryer (long term purchase) + monthly essentials (soap / handwash), it shows up as one single item spent on amazon!

Thus, there is a need for an intelligent middleman for tracking your finances better.

What it does

aurum.ai deeply integrates with all of your banking information - beyond what your API actually provides you through intelligent email parsing. Every transaction based system involving money typically send out emails regarding what actually happened. On a fixed schedule, aurum does the following:

• Pull all unread emails that look like they're a finance based app, and imports them by parsing the full HTML from the email using an LLM.
• Parses existing transactions and recent transactions, and updates key learnings from each week regarding spending habits

• After each workflow run, it creates a set of "pings" that the user has to act on - either to proactively gain information/rules in the beginning, or point out that something about their spending is anamolous

• Aurum is also deeply integrated with the actual database containing the information, and is LLM-first. This means that it can generate charts and insights that are hypepersonalized to the user!

How we built it

• 100% vibe coded, including setup of google dev accounts using browser use.
• Mastra for orchestrating agents
• Created a simulated human to step through how the app helps a person over the course of a year.

Challenges we ran into

• Understanding what to build first - I ended up building a lot of API based infrastructure to connect bank accounts with plaid, actually pull from emails, and integrate LLM based tools. However, formulating the idea to actually use agents/workflows to create insights and improve over time was quite hard. Thus, I got late at actually building this.
• Vibe coding spirals and messes up sometimes! Also running multiple agents to iterate on the same codebase is quite complicated.

What we learned

• There is a gap in the market for an intelligence layer for personal finance.
• Agents can do 90% of the heavylifting. the last 10% is still challenging.
• Vibe coding can lead to the next 5 member billion dollar company. I know nothing about how the product was written, but it is functional. Probably wont scale, but we learn as we go.

What's next for aurum.ai

• Build this out as a personal project to actually track my finances really well. Once thats done, ship it and sell.

Built With

• ag-ui
• mastra
• react