The author presents IntelShed, an open-source system for open-source intelligence (OSINT) that aggregates 50 public data feeds to produce automated 24-hour security briefings. The architecture integrates hybrid retrieval, GNN-based entity resolution, federated learning with differential privacy, and LLM-compiled multi-agent orchestration.

  • Hybrid retrieval uses pgvector HNSW, tsvector GIN, reciprocal rank fusion, and a BGE cross-encoder reranker on ~2,000 indexed chunks in PostgreSQL 16.
  • Entity resolution employs the FollowTheMoney schema with ~142k entities and 173k edges, utilizing Splink for probabilistic matching and GNN-predicted links.
  • The GNN pipeline includes GAT for correlation, a T-GCN forecaster for 24-hour predictions, and GNNExplainer for attribution.
  • Federated learning implements FedAvg with DP-SGD, currently tested on a single GPU node but planned for two-node exchange.
  • A trust layer applies two-sided Bayesian source adjudication with corroboration caps, reducing the rate from 100% to 16.8%.
  • Multi-agent orchestration features an 8-phase workflow with loop detection and a "Vibe Graphing" module that compiles natural-language queries into execution graphs.

The author seeks community feedback and contributions, specifically requesting independent benchmarks for retrieval quality, security reviews of the MCP tool surface, validation of federated learning with multi-node gradient exchange, and statistical evaluation of the Bayesian trust scoring methodology.