Perspectivas de Desenvolvimento de Assistentes Conversacionais com LLMs: Um Estudo de Caso

Abstract

Conversational assistants based on Large Language Models (LLMs) have become increasingly common, but designing them in a reliable and maintainable way is not straightforward, as it involves architectural decisions across multiple dimensions. This work reports the iterative development of Finbot, a conversational assistant for personal finance management integrated with Telegram, built through three progressive versions: a monolithic system, a multi-workflow architecture with LLM-based entity extraction, and a multi-agent architecture with atomic tools. The versions were compared using qualitative software engineering criteria and quantitative metrics collected with the N8N Evaluations module under the LLM-as-Judge paradigm. Finally, the work discusses the gains and trade-offs of each version and constitutes a practical report on the development of an LLM-based conversational assistant, from prototype to final version, which may serve as a reference for similar projects.