AI Engineer World's Fair 2026 — unofficial import demo

Every enterprise AI team faces the same strategic question: where in the stack should a small team focus its effort? Models, frontends, and agent frameworks evolve rapidly and are increasingly commoditized. But regardless of how these layers mature, AI in enterprise settings remains bottlenecked by the same underlying problem: structured data is siloed across systems of record with domain-specific schemas, and the unstructured data needed to contextualize it sits in entirely separate systems, with its own systematic complexities. The durable work is cleaning, curating, and semantically modeling this data in an AI-first manner so that any client — chat, workflow, or otherwise — can query across it. That's the moat. At the Gates Foundation, my team built and deployed our foundation-wide knowledge graph on Neo4j that unifies structured and unstructured data behind a single MCP server. The graph itself is modeled for agentic consumption: natural hierarchies are projected as traversable paths rather than flattened tables, and unstructured documents are semantically chunked, tagged, and mapped to structured entities at ingestion time using AI-driven ETL. The result is a semantic layer where an agent can express a complex cross-system question as a concise graph query and receive an accurate answer. This talk is an architectural walkthrough covering the end-to-end pipeline: AI-based extraction and semantic chunking of unstructured documents, the agent-first data modeling decisions, design considerations for our MCP server, and how we handle graph-based retrieval evals. We'll walk through real query sessions showing Claude interacting with the graph through both chat and workflow integrations. The intended takeaway is a practical framework for where a small enterprise team's investment compounds — and why that investment is the data model, not the layers above it.