Virtual Engineer

AI-assisted chiller troubleshooting demo powered by MongoDB Atlas. A field engineer describes a symptom; an LLM agent retrieves operational facts and knowledge through MCP tools, with live visibility into MongoDB retrieval patterns in the demo UI.

What it demonstrates

• Deterministic grounding — asset resolution, alarms, telemetry, service history

• Probabilistic retrieval — Atlas Vector Search and hybrid case search

• Real-time transparency — query patterns, pipelines, and latency in the UI

• Two demo paradigms — Evidence Board (technical) and Field Chat with X-ray (business)

Related MCP server: Florentine.ai MCP Server

Architecture

Demo UI (React)  →  LLM Agent (SSE)  →  MCP Server  →  Atlas
                                              ├── Database (operational data)
                                              ├── Vector Search (knowledge, cases)
                                              └── Atlas Search (case notes)

See docs/architecture.md for details.

Prerequisites

• Node.js 18+

• MongoDB Atlas cluster

• OpenAI or Anthropic API key (for chat), or MongoDB Grove key (MDB_GROVE_API_KEY) for internal demos

Quick start

cp .env.example .env   # add MONGODB_URI and an LLM key (OPENAI_API_KEY, ANTHROPIC_API_KEY, or MDB_GROVE_API_KEY)

npm install
npm run seed:drop      # load sample data into virtual_engineer database

# Terminal 1 — backend (MCP + chat API)
npm run mcp:dev

# Terminal 2 — demo UI
cd frontend && npm install && npm run dev

Open http://localhost:5173.

Demo scenarios

Starter prompts are on the Overview tab. See scripts/data/README.md for the full scenario matrix.

Scripts

Documentation

License

MIT

Requirements

Product and design rationale: Virtual Engineer High-Level Requirements and Design Logic.md