Disclosure Signal Research Engine
Research system that turns public financial disclosures (legislator trade reports, institutional holdings, insider filings) into traceable signals, monitors a portfolio and delivers a cited daily digest. A person makes every trading decision.
Architecture
Python engine with a Typer CLI that parses filings, including scanned PDFs through an OCR pipeline, into normalized signals linked to their source. A deterministic scanner handles revaluation, technical levels, alert deduplication and alert lifecycle; scheduled LLM jobs only write the narrative, and citations are attached in code. A spending cap and budget ledger degrade the jobs at 70% and 85% of the monthly budget, and a gate skips the model when nothing changed. Runs daily on EC2 with a React dashboard and Telegram delivery.
Highlights
- In daily production use
- The code decides, the model writes
- Monthly LLM budget with graceful degradation
- 300+ offline tests with replay fixtures
- Advisory only, never places trades