Intelligent Multilingual Financial Communication Systems: Deep Learning Approaches For Code-switching Optimization In Kenyan Digital Finance

Abstract

This thesis presents the design, development, and evaluation of an AI-driven multilingual financial education system for the Kenyan market, addressing the critical gap between financial literacy and financial access in a digitally connected but underserved population. The system employs a Retrieval-Augmented Generation architecture, combining Anthropic's Claude Sonnet 4 with a FAISS vector index over 86 curated chunks spanning SACCOs, Treasury Bills, Money Market Funds, NSE equities, pension schemes, and personal budgeting. A dual-layer code-switching pipeline classifies user input as English, Kiswahili, or mixed with 99.4% accuracy, enabling natural multilingual interaction without forcing users into a single-language paradigm. The full-stack application is deployed on Vercel and Render with PostgreSQL persistence, and incorporates JWT authentication, role-based access control, a three-tier memory architecture, and a confidence scoring framework that communicates uncertainty transparently to users. Evaluation yields an MRR@5 of 0.87, a mean factual accuracy of 4.45/5.00, and a user satisfaction score of 3.95/5.00 across a 20-participant study, with a mean response latency of 2.1 seconds on the live deployed system. The results demonstrate that RAG-augmented large language models, combined with domain-specific engineering and accessible cloud infrastructure, can deliver production-quality multilingual financial education to communities that have historically lacked formal financial guidance.