Beyond LLM Wrappers: Context Engineering and the Three-Engine Architecture for Reliable Financial AI
Three engines handle finance: knowledge graph, deterministic optimizer, multi-agent LLM layer. Over 15 deployments, observed hallucination fell from about 7-9% to below 1%.
The claim
Three engines plus context engineering keep financial AI answers exact and traceable.
The problem
Financial answers must be accurate, traceable, and auditable.
Language models cannot guarantee calculations or grounded recommendations alone.
Single-agent pipelines invent invoices and unsupported claims.
How it works
A financial knowledge graph stores relationships, time, and provenance.
A deterministic optimization engine computes numbers and decisions.
A multi-agent LLM layer interprets intent and validates claims.
Context engineering selects and validates context across ERP systems.
Reusable mappings and templates speed new deployments.
Figures
What changes
Across 15 deployments, validation cut observed hallucinations from 7-9% to below 1%.
Deployment time fell from about six weeks to two.
The findings are observational, not causal.
Limits
The study observes deployments; it runs no controlled experiment.
ERP data quality and tenant schemas stay hard problems.
Cross-tenant generalization and privacy remain open challenges.
Source · arXiv