Midas

Landing Page
Midas v1
The Oracle Prediction with Parameters
Dashboard with Predictions
Took a minute to plan Midas' brain
11 Red bulls in one night...
Three.js started working!
"Why isn't the AI working?"
Knocked out.

Inspiration

Our friend Josh had a problem buying Pokémon cards and in-game currency (Valorant). So much that he ended up in an unpayable $2000 dollar credit card debt. When Josh got into a car accident, he had no way to pay for the damages. Thankfully, he had people around him who helped.

We made Midas to help people like Josh avoid unfortunate circumstances that happen due to overspending. Utilizing Gemini, predictive LSTM, and clear visuals, Midas utilizes modern technologies to solve an old problem: spending too much.

What it does

Midas is a personal financial assistant that aggregates bank transaction data from APIs, categorizes spending, and visualizes your financial landscape on an interactive dashboard. It leverages machine learning to forecast future spending, detect anomalies, and simulate future patterns based on parameters such as reducing spending in one category by 20%. With robust security features like in-rest security features, Midas ensures your financial data remains trustworthy and secure in real time.

How we built it

📊 Data Integration

We integrated Plaid to pull transaction data directly into our secure Convex DB. We used mock data provided by Plaid for this hackathon.

🖥️ Interactive Dashboard

Our responsive shadcn/tailwind css web interface displays intuitive visuals like a monthly cumulative graph with predicted progress and spending based on common categories like groceries.

🔮 The Oracle

We developed an predictive LSTM model that allows users to adjust parameters such as:

🍎 Spend $500 less on groceries this month.

🔔 Spend $20 on a new subscription this month.

👕 Stop spending on clothes for this month.

And receive the predicted path of their spending for the month.

🚨 Anomaly Algorithm

We implemented a statistical algorithm using interquartile range inference to send real time email alerts to users when suspicious transactions happen or budget limits are crossed.

Challenges we ran into

No hackathon is without challenges, and this golden dream wasn't an exception. Our primary headaches were:

Getting reliable financial data for the training of the predictive model Individual financial data is not a readily available resource, and this made training of our model a lot harder. We got past this hurdle by building simulator that simulates the financial activities of a person for 2 years.
Ensuring data security in our application Being a first time financial application for our team, Midas's security was a big feature we wanted to implement as we knew our previous projects' security measures were not enough. To address this, we tried to store an audit log of all transactions to and from the database on the Midnight blockchain. Unfortunately, this fell through due to time constraints. We overcame this challenge by incorporating data encryption in-rest and deploying our fullstack application with https.

Accomplishments that we're proud of

"I'm proud we lost the last 3 hackathons and still went all in for this one." - Anirudh Kuppili | Project Manager | Management and Direction

"I'm proud that we were able to adapt to challenging situations" - Isaac Alazar | Financial Techy (Bizdev) | Plaid, Live Demo Lead

"Proud we built an end-to-end ML System overnight. Hope we win and make it worth." - Vikram Harikrishnan | AI/ML Engineer | The Oracle

"I am very proud to work with experts that achieve their goals no matter what" - Arslan Kamchybekov | Fullstack Developer | LLM, Charts, Graphs, Website

"I am extremely proud that we tried new things and explored new tech" - Joshua Jung | Data Expert and Creative Lead | Anomaly Algorithm and Video Skit Director