I developed an LLM-powered personal finance assistant integrated into Telegram, enabling users to log expenses and income through natural conversation. The core intelligence of the system is implemented using a LangChain-based agent, which interprets user intent, extracts structured financial information, and decides the appropriate action through tool invocation. The agent supports both text inputs and receipt images, allowing seamless multimodal interaction.

Key Contributions:

• Designed and implemented a LangChain agent with structured reasoning flow for financial logging and querying.
• Integrated tool calling for actions such as transaction creation, categorization, and database insertion.
• Implemented short-term conversational memory to maintain context across multi-turn interactions.
• Developed robust prompt design to ensure consistent extraction of amount, category, and description from noisy user inputs.
• Enabled multimodal understanding by combining text-based reasoning with receipt image interpretation.

Technical Stack:

• LangChain / LangGraph for agent orchestration and state management
• Gemini 2.5 Flash as the reasoning and multimodal LLM
• PostgreSQL for persistent transaction storage
• Telegram Bot API as the conversational interface

Outcome:

This project demonstrates my ability to design and implement a production-oriented LLM agent that goes beyond simple prompt-response patterns. It highlights my experience in agent reasoning, tool-based execution, memory handling, and real-world integration with databases and messaging platforms.

GitHub: Github Project