Rocket League - AI Powered Improvement

Summary

I've had an experience in Rocket League where I will play relatively well for a while, getting to 1580 MMR (GC1 Div 4), then a week later I fall to 1250 and play very poorly for a while. It suddenly occurred to me that I have access to at least a few hundred old replays through ballchasing.com, maybe there's some way to analyze the replays and identify exactly what I'm doing differently when I play well vs poorly?

Getting Started

To get an idea of the maturity of tools in the ecosystem, I first just opened up Cursor and in so many words said: "Build a tool that pulls all my replays from ballchasing.com, analyzes them, and figures out what I'm doing differently in wins vs losses." Based on my subjective experience, I suggested some places to look, for example:

• Which direction I face when I jump - Maybe I'm always slightly facing the wrong way?
• Actions per minute - Maybe when I'm playing well I'm pressing more or less keys?
• Hesitation - Maybe I overthink when I'm playing bad and jump late?
• Powerslide or boost key presses - Maybe rapidly tapping powerslide or boost is an indicator that I'm not confident in timings?
• Boost usage and collection - Maybe I use or collect boost in different patterns?
• Recovery decisions - Maybe I don't recover as fast or path wrong?

To my surprise, Cursor quickly identified ballchasing.com's API, instructed me on how to get an API key and began pulling all of my replays. Great!

After that I iterated with Cursor and began testing out theories using a Python package called boxcars.

Moving to an App

Eventually I realized that what I might actually be looking for maybe not be achievable by just having cursor review replays and test how much boost I had or how quick I recovered. I realized being better in a team game may actually require built in expert knowledge. I spent a while building and encoding expert advice into detectable patterns.