An AI business analyst for hotel management. It combines quantitative booking data (counts, revenue, occupancy) with qualitative context from internal meeting notes to answer questions like "Why did Deluxe room revenue drop this summer?" or "Where should we run our next promotion?" — citing the actual numbers and the specific meeting/decision behind them.
Two ways in:
- Chat UI — a simple web chat for hotel managers.
- MCP server — the same tools exposed over Streamable HTTP for any MCP-compatible client (Claude Desktop, etc).
Data sources are swappable: local mock fixtures for quick local dev, or a live Google Drive folder (Google Docs for meeting notes, a Google Sheet for bookings) kept in sync automatically.
chat-ui (Next.js/Bun) ──/api/chat──► mcp-rag-engine (FastAPI/LangGraph, Gemini) ──► Qdrant
│
├─ /mcp (Streamable HTTP, external MCP clients)
└─ Google Drive sync (meeting notes + bookings)
| Component | Stack |
|---|---|
mcp-engine/ |
Python 3.12, FastAPI, LangGraph agent (Gemini), mcp SDK, Qdrant, uv |
chat-ui/ |
Bun, Next.js (App Router), TypeScript |
vector-db |
Qdrant (Docker) |
# Backend
cd mcp-engine
uv sync
cp .env.example .env # fill in GEMINI_API_KEY
docker compose -f ../docker-compose.yml up -d vector-db
uv run python scripts/generate_mock_data.py
uv run python -m app.rag.ingest
uv run uvicorn app.main:app --reload --port 8080
# Chat UI (separate terminal)
cd chat-ui
bun install
cp .env.local.example .env.local
bun run dev # http://localhost:3000Or run everything in Docker: docker compose up -d --build from the project root.
To use real Google Drive data instead of mock fixtures, set DATA_SOURCE=gdrive,
GOOGLE_SERVICE_ACCOUNT_PATH, and GOOGLE_DRIVE_FOLDER_ID in mcp-engine/.env.
Infrastructure is defined in Terraform (deploy/terraform/) and provisions an ephemeral AWS
Lightsail VM that pulls pre-built images from ECR — good for spinning up a demo, then tearing
it down.
aws configure # or export AWS_PROFILE=<your-profile>
make deploy # build + push images to ECR, then terraform apply
make output # show the chat UI URL
make destroy # tear everything down — the only way to stop billingmake deploy PROFILE=<name> to use a specific AWS CLI profile. See CLAUDE.md for
implementation details and known gotchas.