Selected work

Relationship intelligence with traceable graph reasoning

NYIC leadership experience linking entities, business relationships, graph paths, and documentary evidence for institutional research.

Engineering lessons

When does a graph relationship become usable research evidence?

  • Resolve entities before relying on paths between them.
  • Connect graph traversal with supporting documents and validated extraction.
  • Assess identity, path relevance, and evidential support as separate questions.

Context and professional contribution

Institutional relationship research asks questions that span multiple records. A company may connect to a sponsor through a transaction, to an executive through a prior role, or to another organization through ownership. Relevant evidence can be distributed across enterprise records, filings, research documents, and relationship-management activity.

This study describes NYIC leadership experience designing and implementing a knowledge-graph platform for relationship intelligence. The work covered the canonical entity model, entity resolution, graph-based research workflows, document enrichment, tool interfaces, deployment, and evaluation. The public account keeps the institution and confidential relationship information anonymous.

The engineering objective was to make relationship paths useful and inspectable. A returned connection needed to carry the evidence behind its constituent relationships, including uncertainty where identity or support remained unresolved.

Establish a canonical model of entities and relationships

The implementation linked companies, sponsors, executives, transactions, and customer-relationship-management activity across enterprise and external sources. Entity resolution, provenance tracking, and confidence scoring were part of the model. These capabilities made ownership and business relationships traceable to their sources.

Canonical identity is a foundation for graph reasoning. A company can appear under several names, and similar names can refer to different organizations. An incorrect merge can create convincing paths that never existed. Preserving source identifiers and confidence allows uncertain matches to remain reviewable before they influence broader research.

Relationships also need context. Ownership, employment, sponsorship, and participation in a transaction express different meanings. A path explanation becomes useful when the reader can understand each edge and the evidence supporting it. A generic connection score alone would conceal the structure that makes the relationship relevant.

Combine graph traversal with document retrieval

GraphRAG workflows ran on Amazon Bedrock AgentCore Runtime. Graph traversal identified relationship paths, while vector retrieval supplied supporting documents. Typed Python and FastAPI tools supported target discovery, ownership tracing, relationship lookup, and path explanation. Their responses included both paths and source evidence.

This architecture connected structural reasoning with source inspection. Traversal could identify a candidate chain of relationships, and retrieval could provide the passages needed to examine those links. The model then had a bounded evidence package for explaining the path in language a researcher could review.

The design also made uncertainty localizable. A path may be structurally valid while one edge rests on weak or ambiguous evidence. Preserving support at the edge level allows a reviewer to distinguish a well-established connection from one that needs further diligence.

Enrich the graph through validated extraction

A Snowflake Cortex enrichment pipeline used AI_EXTRACT to identify candidate entities and relationships in filings and research documents. Extracted records were validated against canonical entity identifiers before graph ingestion. Cortex Search retrieved supporting passages by entity and document metadata for path explanations.

The validation boundary was important. Extraction produces candidates that can contain ambiguous names or an incorrectly inferred relationship. Checking against canonical records before ingestion reduces the chance that one unsupported extraction becomes a reusable fact across many subsequent questions.

The passage-retrieval layer also helped preserve the distinction between the structured graph and its source material. A graph edge is a useful representation, while the original document supplies context that may qualify the relationship. A researcher needs access to both when a connection influences a consequential decision.

Expose research tools with scoped access

Research APIs were exposed as Model Context Protocol tools through AgentCore Gateway. Scoped access controlled the underlying services. Reusable AWS Cloud Development Kit constructs deployed the agent runtime, gateway, identity roles, and logging configuration.

The platform integrated governed Snowflake and Databricks information with the agent services. This linked the research experience to enterprise data controls and provided explicit service boundaries for retrieval and graph operations. Tool schemas made the requested entity and operation available for validation.

Relationship knowledge can itself be sensitive. An enterprise connection inferred from restricted records should not become visible merely because it has been converted into a graph path. The architecture’s permission boundaries are therefore part of the meaning and permitted use of the returned evidence.

Evaluate identity, paths, and support separately

The documented implementation included evaluation checks for entity matching, path validity, source attribution, and output schemas. Uncertain identities and unsupported relationships were routed to human review. These checks addressed distinct failure modes that a single answer-quality score would struggle to explain.

Entity matching asks whether the correct records were connected. Path validity asks whether the graph operation returned a coherent sequence. Source attribution asks whether the evidence supports the relationships described. Output validation ensures the result remains usable by the application and its review interface.

Further assessment of this class of system should consider temporal validity and incomplete coverage. A historical relationship may no longer hold, and an absent edge may reflect missing information. These questions help define the evidence a reviewer needs before relying on a returned relationship.

Documented result and transferable lesson

The delivered platform connected canonical entities, relationship paths, validated document enrichment, and supporting passages through a controlled research interface. Review controls preserved a path for resolving uncertain identity and unsupported links before researchers relied on them in a broader analysis.

The central lesson is that graph reasoning becomes useful when the path remains connected to its evidence. The platform’s architecture made that connection explicit and preserved a role for human review where identity or source support could not be established confidently. The same approach can inform diligence, ownership research, expertise discovery, and other enterprise questions whose answer depends on a chain of relationships.