Designing and building a multi-layer agentic AI system for trading signal generation — transforming unstructured data (news, filings, earnings, macro indicators) into structured buy/sell signals through specialized LLM agents with explicit reasoning, memory, and guardrails. Perception, Cognition, and Action layers implemented; Memory layer in active development.
A modular agent architecture that decomposes financial decision-making into specialized layers — each responsible for a distinct cognitive function. Rather than a monolithic model that maps raw text to signals, the system uses explicit agent decomposition with structured state passing, tool use, and a novel Thesis→Antithesis→Synthesis reasoning pattern.
The core insight: financial information is heterogeneous (news vs. filings vs. price action), arrives at different timescales, and often contradicts itself. Agentic decomposition provides a better inductive bias for this than end-to-end prediction.
Tech stack: Python · LangGraph · Qdrant · PostgreSQL/pgvector · Pydantic · OpenAI / Anthropic / Gemini APIs · Jinja2
https://sh1319.github.io/diagrams/agentic_architecture_diagram.html
Specialized agents parse raw data sources into structured analyses:
| Agent | Input | Output |
|---|---|---|
| Price | OHLCV data + chart images | Technical analysis (RSI, MACD, Bollinger) |
| News | Financial news, social media | Sentiment analysis with relevance reranking |
| Filings | SEC filings, earnings calls | Key financial factor extraction via RAG |
| Macro | Economic indicators | Macro-to-sector-to-company chain of thought |
Each agent outputs Pydantic-typed structured objects — not free-form text — ensuring type safety and downstream composability.
Retrieval layer that grounds current analysis in historical context: