RedShift

Speaking to Jump Trading, we figured out there was a huge demand for creative solutions to analyzing market trading strategies.

Our solution works like an AI bully bot: it makes slightly faster prediction (on the scale of nanoseconds) that will show you the faults in your marketing strategy by picking you off the market.

Using an Arduino UNO Q and a Raspberry Pi 4, we simulated a live market driven by real-time Bitcoin data. On top of it, we built an AI “predator” that trades against your market-making strategy, exploiting stale quotes the way a faster, better-informed trader would. You can adjust the predator’s latency and choose from different strategies to test. After each run, an LLM hosted on Vultr writes a security report explaining where your strategy leaked money and how to fix it. We built the project with Claude Code: we wrote a detailed spec of the system, and Claude Code turned it into working code phase by phase, from the market-data recorder and exchange simulator to the dashboard and the Arduino firmware.

We had to overcome issues with our hardware, ranging from soldering for the first time, figuring out all the connections, and syncing the data between all three machines on a private network.

Our working prototype shows detailed feedback on the user's market-making strategy and passed rigorous rounds of testing. Despite all the trouble, we managed to set up and connect all the hardware and get it running the way we needed it. Our hardware communicates with each other and our prototype in a way that provides working real time data.

We learned a lot about hardware, from soldering to writing images to SD cards and hosting several machines on a network to debugging issues with several small components on our breadboard. We also learned how to train and draw useful information from AI models and automate menial tasks with agentic AI.

With more money and time, RedShift can become a very powerful tool for evaluating market-making strategies. A stronger AI model can find and exploit many more subtle and instance-based weaknesses. With better hardware, we could better model the latency between routes of data and evaluate our strategies in more extreme scenarios. Our interface can also become something useful for individuals and beginners in quant trading, both those seeking to learn and those seeking to improve their own strategies.