SleepyQuant: A 12-Agent Crypto Quant Running on a Single Mac
SleepyQuant is a solo experiment running 12 local AI agents on an Apple M1 Max to coordinate a paper crypto trading book. The author shares the technical stack, design choices, and key features of this local-first AI system.
Why it matters
This project demonstrates a novel approach to building a local-first AI system for crypto trading, with insights into agent-based architecture and resource management on Apple M1 hardware.
Key Points
- 112 specialized AI agents coordinating a paper crypto trading book on a single Mac
- 2Tight risk management with small initial capital, post-mortems on losing trades
- 3Agents have focused system prompts and skill handlers, with a dispatcher routing requests
- 4Exploring tradeoffs between a single larger agent vs. 12 smaller agents
Details
SleepyQuant is a solo project that runs 12 local AI agents on an Apple M1 Max machine to manage a paper crypto trading book. The technical stack includes MLX Qwen 2.5 32B Q8 as the primary agent model, DeepSeek R1 14B Q8 for reasoning tasks, and a FastAPI backend with a SwiftUI macOS app. The paper trading book starts with just $78 equivalent, with tight stop-loss and take-profit limits, and a hard daily drawdown limit. Each losing trade is analyzed, and the failure vault is published weekly. The 12 agents have specialized roles like trading lead, futures/spot executors, CFO, CTO, and more, with a dispatcher routing requests to the appropriate agent. The author is exploring the tradeoffs between a single larger agent versus the 12-agent decomposition approach, as well as efficient unloading strategies for the LLM models on Apple Silicon.
No comments yet
Be the first to comment