Scholarlink

This paper presents a conversational expense tracking system accessible through WhatsApp messages. Users can log expenses or query spending habits using natural language, eliminating the need for specialized finance apps or spreadsheets. The system employs a hybrid LLM approach, routing simpler queries to a lightweight model for speed and cost efficiency, while complex inquiries are handled by a more capable model. Built with Flask, LangGraph, and Twilio, our system demonstrated faster response times than single-model alternatives without sacrificing accuracy. The method accommodates informal interaction in users’ current messaging practices instead of necessitating context switching to dedicated apps. Our results illustrate the strength of hybrid model routing in real-world AI applications, elucidating how smaller and larger models can cooperate to build scalable, inexpensive, and efficient solutions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Conversational Expense Tracking Using a Hybrid Language Model Pipeline

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Prakhar Shukla,
  • Om Dabral,
  • Jaspreet Singh,
  • Hardik Sharma,
  • Samriddhi Dwivedi,
  • Shivanshu Sahu,
  • Bagesh Kumar

摘要

This paper presents a conversational expense tracking system accessible through WhatsApp messages. Users can log expenses or query spending habits using natural language, eliminating the need for specialized finance apps or spreadsheets. The system employs a hybrid LLM approach, routing simpler queries to a lightweight model for speed and cost efficiency, while complex inquiries are handled by a more capable model. Built with Flask, LangGraph, and Twilio, our system demonstrated faster response times than single-model alternatives without sacrificing accuracy. The method accommodates informal interaction in users’ current messaging practices instead of necessitating context switching to dedicated apps. Our results illustrate the strength of hybrid model routing in real-world AI applications, elucidating how smaller and larger models can cooperate to build scalable, inexpensive, and efficient solutions.