Regel‑konsistent regnskaps‑LLM – domene‑pretrening, RLHF og finjustering for bokføring

Offentlig prosjektsammendrag

ReAI AS will develop an explainable, rule-consistent AI accounting engine for Norwegian SMEs. Current tools rely on static rules or generic ML and struggle with SAF-T and VAT updates, transparency, and data privacy. * Primary objective: Build and validate a domain-specialized LLM that classifies and explains bookkeeping entries with fewer than 1 rule violation per 1,000 transactions and passes SAF-T export validation. * Key challenges: domain adaptation, guaranteed rule compliance and calibrated confidence, auditor ready explanations, privacy across clients, handling regulatory drift. * Method: domain-adaptive pre-training, rule-constrained RLHF with automatic validator, explanation generator linked to accounts and rules, federated learning with differential privacy, offline tests and blinded auditor review. * Expected result: Prototype that automates bookkeeping and reconciliation with verifiable compliance and privacy, ready for integration into ReAI’s system.