Two Graphs, One Agent: From Process PDFs to Auditable Workflows

I built a system that turns scattered Quality Management System documentation into business processes that an AI agent can inspect, traverse, and execute.

The first graph is a knowledge graph. It extracts processes, steps, decisions, roles, documents, business entities, database tables, and stored procedures from PDF and DOCX files. The second is an execution graph: it turns selected processes into conditional workflows with deterministic rules, LLM-assisted classification, human-approval gates, and audit trails.
An initial prototype generates and analyzes a process graph from a document. The operational version adds a typed ontology, SQL Server grounding, semantic retrieval, persistent snapshots, extraction caching, and manual overrides. Its current snapshot contains 2,572 nodes and 6,330 relationships.
Live, I will ingest a sanitized process document, inspect the resulting subgraph, trace a process to its actors and data dependencies, and run a controlled workflow that exposes its execution path and stops for human review before a critical action. I will show the working system, code, graph data, and traces—no slides.