ExpenseSense: Privacy-Preserving On-Device Personal Finance Tool Calling with Sub-2B LLMs

For personal finance, cloud-based inference poses privacy risks, highlighting the need for local on-device execution. Yet, the tool-calling performance of sub-2B small language models (SLMs) remains under-explored. We introduce Expensense, a privacy-first personal finance assistant, and benchmark on-device tool calling using a 115-question test suite. We compare single-agent (joint routing and parameter extraction) against dual-agent (router-specialist) pipelines across six local models (0.8B-2B). Dual-agent decomposition improves correctness (e.g., Qwen3.5-2B Correct Ratio improves from 60.2% to 77.6%), as well as decreases average latency by up to 48% (e.g., from 12.0~s to 6.2~s for Qwen3.5-2B). It also shifts errors toward safer failure modes. Qwen3.5-0.8B dual achieves 89.1% Task Accuracy with 3.0~s latency and 1.2GB RAM, demonstrating the viability of fully private, on-device financial assistants. All code, data, and prompts are available on GitHub.