Dodge AI — Order to Cash

Problem

SAP Order-to-Cash data is fragmented across tables, making it difficult to visualize document flows and query relationships in plain English.

Architecture

React + Vite frontend for graph viz and chat; FastAPI backend orchestrating Gemini LLM calls, PostgreSQL for relational storage, and NetworkX for in-memory graph building.

Challenges

• — Mapping relational foreign keys to graph edges efficiently
• — Ensuring LLM generates correct SQL without hallucinating
• — Preventing expensive or unbounded database queries

Engineering Decisions

• — Chose PostgreSQL over Neo4j to leverage LLM SQL proficiency
• — Injected schema and used CoT prompting for accurate SQL generation
• — Implemented read-only SQL execution and statement timeouts

Tradeoffs

• — LLM inference latency vs accuracy
• — Relational database vs native graph DB for graph-like traversal

Future Work

• — Support for more complex graph algorithms
• — Real-time sync with SAP systems