MARA - Multi-Agent Research Assistant

Automated research platform with 4 specialized agents orchestrated by LangGraph for comprehensive research workflows.

Role: AI/ML Engineer. Team: Individual. Status: Complete. Year: 2025.

Overview

MARA (Multi-Agent Research Assistant) automates the research process using 4 specialized AI agents orchestrated by LangGraph:

1. Research Agent: Gathers information from multiple sources 2. Summarization Agent: Condenses findings into key insights 3. Citation Verification Agent: Validates sources and fact-checks claims 4. Report Generation Agent: Produces structured research reports

Architecture:

  • Orchestration: LangGraph for agent workflow management
  • AI: Vertex AI Gemini backbone for all agents
  • Memory: ChromaDB for vector storage, DuckDB for metadata
  • Frontend: React dashboard for research visualization

Why This Architecture:

  • LangGraph enables complex agent handoffs and parallel execution
  • Specialized agents outperform single-model approaches on research tasks
  • ChromaDB + DuckDB hybrid provides fast retrieval with rich metadata
  • Vertex AI provides consistent performance across all agents

Key features

  • 4 specialized research agents with distinct capabilities
  • LangGraph orchestration for complex workflows
  • Vertex AI Gemini backbone for consistent performance
  • ChromaDB vector memory for semantic search
  • DuckDB metadata storage for structured queries
  • Citation verification with source validation
  • Structured report generation with formatting
  • React dashboard for research visualization

Technical details

  • Orchestration: LangGraph with custom state management
  • AI: Vertex AI Gemini 2.0 Flash for all agents
  • Vector Store: ChromaDB with sentence-transformers embeddings
  • Metadata: DuckDB for fast analytical queries
  • Backend: FastAPI with async processing
  • Frontend: React with research workflow visualization
  • Deployment: Cloud Run with min-instances=0

Challenges

  • Designing agent handoffs that preserve context across workflow stages
  • Balancing agent specialization vs generalization
  • Implementing reliable citation verification without hallucination
  • Managing memory across long-running research sessions

Tags: Python, FastAPI, LangGraph, Vertex AI, ChromaDB, DuckDB