Executive Track Record Engine

The problem

Investment teams need to know what an executive actually led, and when. Proxy statements (DEF 14A filings) and change-of-officer notices (8-K Item 5.02 filings) hold that history, but standard search by meaning similarity will happily attach a Microsoft project to an Apple question because the name embedding is close.

The knowledge graph, a database that stores who worked where and when as connected facts rather than loose text, is the source of truth for people, companies, roles, and tenures. Vector search only supplies supporting text. The language model narrates the pack; it does not invent employers or dates.

How it works

1. Real SEC EDGAR corpus: fixed 10 major public companies (large-cap), proxy statements as primary source
2. Human-verified reference set: 25 executives, 71 relationships, 8 test queries
3. Local extraction: Qwen3-8B entity and relationship candidates scored vs the reference
4. Dual stores: Neo4j (a graph database) for multi-hop structure; Qdrant (a vector database) for text segments
5. Hybrid retrieval: 1–2 hop neighborhood + stock-symbol-filtered vectors
6. Memo with source citations: AI writes, then a checker rejects claims not tied to a filing

Results

Extraction without fine-tuning: executive entity recall 100%, WORKED_AT (employment relationships) recall 90.6%, HELD_ROLE (specific job titles) recall 87.5%. Initiative extraction remains weak (~33%) and is reported as diagnostic only. Recall here means the fraction of correct facts the system successfully found.