JourneyMeshMulti-Agent AI Travel Planner LangGraph + MCP + LLMOps

Overview

JourneyMesh - every journey, intelligently connected. It is built as an agentic system rather than a chatbot: a supervisor decides which specialists a request actually needs, every external call is authorised before it leaves the process, every draft is measured, and nothing is final until a human approves it.

The whole workflow runs end to end with no third-party credentials: without an API key each provider falls back to a deterministic adapter whose output is labelled ESTIMATE, so nothing is ever presented as a live price unless a provider confirmed it.

Architecture

• Interface: React 18, TypeScript, Vite, TanStack Query, Tailwind CSS, i18next (English, Bengali, Hindi)
• API: FastAPI and Pydantic v2 behind security middleware (request id, size limit, rate limit, headers)
• Orchestration: LangGraph + LangChain with PostgreSQL checkpoints, so the workflow pauses at review and resumes later
• Tools: Model Context Protocol client with aviation, search and custom weather MCP servers
• Storage: PostgreSQL, SQLAlchemy 2.0 and Alembic for trips, results, reviews, conversations and audit events
• Delivery: one multi-stage Docker image, GitHub Actions quality gate → GHCR → OVHcloud VPS behind a shared Caddy proxy

Dynamic supervisor

The supervisor reads the request and returns a routing decision; it never plans the trip itself. A weather question runs one agent, a full journey runs five. A stated budget pulls in flights and hotels, a multi-night itinerary pulls in hotels, and weather is opt-in. Routing is deterministic by default; a configured model may widen or narrow the selection but only within the known agent set.

Specialist agents

• Flight agent - resolves airports and normalises routes, never inventing flight numbers or fares; every price is labelled LIVE, SEARCH_DERIVED, ESTIMATE or UNAVAILABLE.
• Hotel agent - derives a nightly ceiling from budget, trip length and travel style, then ranks candidates on price fit, rating, party size, distance and interests.
• Weather agent - forecast for the travel window turned into packing and activity guidance.
• Budget agent - structured cost picture with per-line provenance and within_budget / near_limit / over_budget status.
• Itinerary agent - deterministic day structure with pacing, rest and indoor swaps when rain or heat is likely.
• Final response agent - assembles the approved slices into one validated journey in the traveller's language.

Human-in-the-loop and selective re-execution

Each specialist owns exactly one slice of a shared TravelState and agents never call each other. When the reviewer requests a change (“find a cheaper hotel but keep my flights”), the supervisor re-runs only the affected agents - hotel, budget, itinerary - and every other slice stays byte-for-byte identical, up to a revision limit.

Guardrails and security

• Input guard: payload size, unsafe markup, weighted prompt-injection rules, unlawful-intent screening, travel relevance and semantic checks.
• PII redaction (passports, IDs, Luhn-checked card numbers, IBANs, emails, phones) before anything reaches a model, MCP server, log or audit trail.
• MCP tool guard on every call: allowlist, calling agent, arguments, risk and budget.
• Output guard: schema validation, completeness, credential and unsafe-URL detection, budget arithmetic and itinerary length.

LLMOps: evaluation and observability

A ten-dimension evaluation module - relevance, completeness, groundedness, consistency, tool correctness, schema validity, safety, language correctness, itinerary feasibility and budget consistency - decides everything decidable by rule and uses an LLM-as-judge only where necessary. Structured JSON logs with PII redaction, in-process tracing and metrics are always on; optional LangSmith tracing shows each graph run, agent, model call and MCP tool call as one nested trace, named by revision.

Quality

218 pytest tests, 21 Vitest + React Testing Library tests and an offline evaluation suite, enforced by a GitHub Actions gate before any release.

Try the live demo · View the source on GitHub