Back to Projects
LangGraphConvNeXtGAFPhi-3.5pgvectorPython
Hybrid-Adaptive Quant Trading System
A production-ready multi-agent trading brain combining Gramian Angular Field (GAF) pattern recognition, ConvNeXt-Tiny neural network for regime/direction/volatility prediction, and LLM-as-a-judge confidence calibration with pgvector memory.
Multi-Agent
Architecture
Phi-3.5-Mini
LLM Backend
ConvNeXt-Tiny
Vision Model
0.78
Confidence Threshold
31K tokens
Context Window
pgvector
Memory Backend
# Methodology
Transforms time series into GAF matrices for pattern recognition. Uses 3-channel GAF (GASF/GADF/heatmap) with ConvNeXt-Tiny for neural pattern classification. Implements Binary Symmetric Channel model for confidence calibration with adaptive refinement loops when confidence < 0.78.
Core Model
# Data Source
Real-time market data via yfinance (Futures, Crypto, ETFs)