Building Luca: An AI Agent for Finance and Accounting Workflows (www.leapfin.com)

Leapfin announced Luca, an AI agent for finance and accounting engineered to eliminate hallucinations, provide full audit trails, and deliver deterministic outputs suitable for SOX/audit requirements. Unlike systems that run raw LLMs as the system of record, Luca uses an “Architect–Builder” separation: Claude (the LLM) acts as a constrained architect that drafts plans and generates code in a custom, heavily restricted JavaScript-based DSL, while a deterministic builder and graph engine execute those plans with provable repeatability. Luca is already in production, automating order-to-cash workflows at scale, handling complex ASC 606 revenue recognition, generating journal entries across contract lifecycles, and performing compliance reviews in minutes instead of weeks. The architecture enforces trust through constraints and a rigorous data foundation: universal data standardization across sources, a Connected Accounting Map (graph of causal links from events to journal entries), event-sourced immutable state, and a bi-temporal data model (business vs. system time). All code passes static analysis that validates semantics, maps affected graph nodes, and produces an execution plan; changes are tested in a sandbox copy of production data before human approval. The result is an auditable, provable system-of-record where AI designs workflows but cannot make nondeterministic or destructive changes—demonstrating a practical blueprint for applying LLMs in highly regulated enterprise domains.