Building a Durable AI Moat Through Knowledge Graphs
AI Engineergo watch the original →
the gist
The Gates Foundation built a defensible AI advantage by modeling 25 years of grant data into a Neo4j knowledge graph, using it as a single source of truth for agents via the Model Context Protocol (MCP).
The Strategic Intelligence Platform
The Gates Foundation developed the Strategic Intelligence Platform (SIP) to solve the problem of siloed operational data across 25 years of grantmaking. Rather than focusing on the UI or the LLM provider, the team built a durable moat by creating a unified knowledge graph in Neo4j. This graph acts as a semantic layer that agents traverse to answer complex organizational questions, ensuring that answers respect historical reporting conventions and internal hierarchies.
Structuring Data for Agentic Retrieval
To make the data agent-ready, the team implemented a rigorous curation pipeline that bridges structured and unstructured data. Key techniques include:
- Semantic Graph Modeling: The team models multiple overlapping hierarchies, such as funding paths (DAGs) and management structures, using derived edges to precompute rollups for complex organizational queries.
- Entity Stitching: By identifying common entities across siloed systems of record, the team created a traversable graph that allows agents to resolve cross-system queries that were previously impossible to answer.
- Unstructured Integration: Documents are ingested, semantically chunked, and tagged as nodes within the graph, allowing for both full-text search and pure graph-based retrieval.
- MCP Implementation: The platform is exposed to Claude via a custom-forked MCP server, which allows the team to pass state, conversation IDs, and message metadata back to the core system.
Governance and Feedback Loops
Because AI increases the accessibility of sensitive data, the team treats governance as a first-class citizen in the pipeline. They mask PII and enforce strict entitlements at the graph level. To maintain accuracy, they run targeted retrieval evals that compare agent outputs against established reporting standards. When the agent fails or returns ambiguous results, the team uses an LLM-as-a-judge to measure stability and pass-at-one rates, using the findings to refine the schema and domain rules in the knowledge graph.