🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Aizen Autonomus Trading Agent
Aizen is a multi-agent options trading system built for the Alpaca AI Trading Agents Hackathon. It runs a LangGraph state machine that coordinates nine specialized agents — regime, direction, volatility, research, options_structure, portfolio, supervisor, risk, and execution — each with a deterministic core and an LLM that explains the decision in natural language. The math never lives in the LLM; the model only narrates it. The signal stack combines three trained XGBoost models (direction classifier, realized-volatility regressor, and an option-level probability-of-profitability classifier) with a news-driven GATv2 graph neural network that learns the dynamic correlation topology across the option chain and underlying assets. When news features are on, the GNN rewires the graph every cycle and injects a topology bias into candidate scoring. Risk is the only authority on whether an order goes out. The deterministic risk engine enforces position caps, loss limits, and a hard floor on probability-of-profitability, plus a direction-mismatch gate that blocks long puts on bullish signals (and vice versa). The supervisor refuses any cycle where agents disagree, where confidence is below threshold, or where a recently-realized loss exists on the symbol. Every decision — including the rejections — is journaled to SQLite with full per-agent reasoning, so the operator can audit the system after the fact. Aizen is currently paper-trading through the Alpaca paper API. Across the recorded cycle log the supervisor has correctly vetoed contradictory option proposals (most recently a long put on a bullish GATv2 signal) that the previous scoring math would have approved. The codebase, training pipeline, and inference engine are open and reproducible.
4 Sep 2026
