Skip to main content
ZR
Toggle navigation
Work

Strong evidence

Investment Research Graph

Private cross-asset research graph with agent-proposed and owner-approved relationships.

Tier
Strong evidence
Pillar
Finance intelligence
Visibility
Sanitized
Status
proprietary
Date
Jun 2026

Case boundary

Private research system with SQLite graph storage, sqlite-vec embeddings, DuckDB and Parquet signals, MCP tools, HTTP sidecar, scheduled refresh jobs, and owner-approved graph mutations.

Claim boundary

Architecture only; no personal positions, recommendations, or trading performance.

Investment Research Graph is a private local research system for mapping relationships across equities, ETFs, futures, options, FX, commodities, macro factors, themes, supply chains, narratives, and filings.

The public case study is intentionally sanitized: the technical work is the graph architecture, data quality pipeline, agent review loop, and approval boundary, not any personal position data.

System

  • SQLite graph database for nodes, edges, pending actions, staging proposals, and security identifiers
  • sqlite-vec embeddings for semantic node lookup and reconciliation
  • DuckDB and Parquet lake for prices, returns, macro series, financials, and analyst inputs
  • Shared Python tools exposed through both MCP stdio and a FastAPI sidecar

Runtime

  • Nightly refresh pipeline ingests prices, macro data, financials, analyst inputs, graph rebuilds, proposals, and briefs
  • LLM proposer writes candidate relationships to staging while a reviewer scores them before promotion
  • Owner-only curation CLI approves, edits, or rejects pending graph mutations
  • Launchd jobs deliver local research briefs and scheduled refresh runs

Proof

  • Sidecar binds to loopback by default and requires explicit auth before remote exposure
  • Programmatic approval is disabled by default; the owner approval path is the curation CLI
  • Tests cover agents, brief delivery, DB durability, graph relationships, sidecar auth, LLM retry behavior, and signal ingestion
  • Data-source failures are isolated per phase so one vendor failure does not corrupt later refresh work