Impact-Site-Verification: 578d421e-1081-463d-918b-ec5e29c5b9db
<Back

AI Agent Decision Support System

Commissioned Work | Forward Deployed Engineer

Overview

Took part in developing an AI agent that argues the same question separately from the supporting and opposing sides. Built on LangGraph, it calls internal-document search and web search only when needed and keeps search scope and memory separate for each customer company.

How It Works

A graph that flows through: load memory → decide whether search is needed → internal-document and web search → supporting and opposing arguments → merge → save memory.

  1. A step that first decides whether search is needed, so unnecessary searches are not called.
  2. The supporting and opposing arguments each read the same context independently and build cited claims in ReAct style, calling search as needed.
  3. LLM reasoning steps are kept apart from steps with fixed behavior, such as reading and writing memory and merging context.

Per-Company Separation and Memory

Search queries carry the customer company's identifier as a filter, so each company searches only its own scope. Conversation history is kept per user and thread, and the oldest messages are dropped once a cap is exceeded. Answers cite source URLs so the evidence can be traced.

Development

The development environment was standardized with a Dev Container; tests were written with pytest, and LLM calls were traced with LangSmith.

Technologies

PythonLangGraphLangChainVertex AI SearchTavilyReActLangSmithpytestDev Container