AI Agent Decision Support System
Commissioned Work
Overview
Designed and developed an AI agent system that provides logical advice from both Pro and Con perspectives to improve the quality of business decision-making. Built on LangGraph/LangChain/FastAPI foundation, achieving a multi-tenant B2B SaaS architecture.
Architecture
Designed an 11-step stateful workflow using LangGraph. Controlled the entire flow from user memory loading -> search necessity judgment -> RAG/Web search -> context integration -> prompt setting -> Pro/Con agent parallel execution -> response merging -> memory saving -> final answer generation through a graph with conditional branching.
Implemented ReAct pattern (Reasoning + Acting), enabling agents to autonomously judge information sufficiency. Executes internal knowledge search via Google Vertex AI Search and web search via Gemini built-in tools as needed, retrieving and integrating optimal information.
Key Features
Pro/Con Dual Agents: Execute independent reasoning from both supportive and critical perspectives on the same context, providing balanced decision support.
User Memory Management: Persist N:N company affiliations, decision history, and user preferences across threads. Implemented conversation history management and token limit control using LangGraph MemorySaver/InMemoryStore.
Multi-tenant Support: Achieved knowledge isolation and access control per company through Company ID-based metadata filtering.
Development & Quality
Built unified development environment with DevContainer (uv, Oh My Zsh, Claude Code/Gemini CLI integration). Established unit and integration tests with pytest/pytest-asyncio/pytest-cov, and introduced tracing/monitoring infrastructure with LangSmith.