Carma AI

Inspiration

We were inspired to make Carma AI after seeing websites like Cars and bids, which cater to a new type of car buyer that seeks to buy enthusiast cars, modify them, and flip them for a profit.

What it does

It shows a user the potential ROI that you could receive from flipping the car. It also goes over potential flaws with the car and the amount you would need to spend to fix up the car, along with any recalls. It pulls data from Autotrader, Cars.com, and Craigslist, all of which contain local inventories of car dealerships and individual salesmen across America.

How we built it

We used Python in VS Code to build this AI agent. We also used vibe coding with Claude code to build a lot of the front-end components. We largely used React to design the front end.

Challenges we ran into

Some challenges we ran into include the Nemotron model and image analyzer, making the analysis with the AI button take over 5 minutes to load. We largely resolved this by reducing the number of API calls made to the nemetron model. This significantly reduced the loading time and made the UI a lot smoother to use.

Accomplishments that we're proud of

Accomplishments that we are proud of include making the agent accurately calculate the potential ROI on a given vehicle. I thought this would be near impossible to implement, given how the prices of enthusiast cars are often not tracked on websites like KBB, which focus on consumer vehicles. Using the Hagerty bull market helped the model analyze the prices of enthusaist cars andd it helped us make it accuretely predict the ROI.

What we learned

We learned that getting four agents to run simultaneously is pretty difficult. Also, we learned to differentiate which tasks need a language model versus which ones should just be Python logic. Another thing we learned is that caching results is super useful, as it helps us not burn API credits on the same car twice.