SEFA: A Review-Gated Mobile Architecture for AI-Assisted Personal Finance Tracking and Expense Intelligence

Personal-finance tasks often require users to reconcile manual records, bank transactions, alerts, and statement documents across separate tools. SEFA brings these activities into a mobile-first, review-gated system that combines transaction tracking, budgeting, read-only bank synchronization, statement-derived transaction ingestion, financial summaries, notifications, and conversational assistance. Its architecture separates the mobile client, backend services, persistent storage, background jobs, and external financial services. Statement-derived transactions are normalized, categorized, checked for duplicates, and held for review before they enter the main ledger; assistant-prepared financial actions likewise require confirmation before persistence. We evaluated the deterministic statement-processing layer with a controlled synthetic benchmark of 300 transactions across six statement conditions. SEFA achieved 3,880/3,900 exact matches across independently scored parser fields (99.49%), 196/250 correct category suggestions for clear-category transactions (78.40%), and a review-required status for 40/50 deliberately ambiguous transactions (80.00% recall). All 11 labelled duplicate occurrences were detected without false positives or false negatives, giving precision, recall, and F1 values of 1.000 within the controlled benchmark. The evaluation covers deterministic Layer-A processing only; it does not assess PDF/OCR extraction, Azure/OpenAI-assisted processing, runtime efficiency, production security, or real-user outcomes. These results support a review-gated design in which reliable deterministic processing is paired with human review when financial interpretation remains uncertain.