NOTE: Effective July 17th 2026, Amadeus (the only free flight pricing API with a generous request allowance) will shut down their self-service APIs. The enterprise plans still come with 2000 free requests, and aren't proibitively expensive assuming you (like EuroTSP will/does) opt to cache results for a likely safe amount of time such that they will still remain accurate (a cache TTL logarithmically scaled to how far out a flight is) when retrieved from the cache.

---

EuroTSP

EuroTSP is a multi-modal European route optimization engine designed to compute the most cost-effective and/or time-efficient travel sequence across multiple cities, weighted based on cost/time preference. The system combines flights and train data and leverages TSP-based constraint solving (via Google OR Tools) to generate structured, bookable itineraries from heterogeneous datasets.

---

Data Flow

• The client calls the API with selected cities, days per city, time weighting, and a trip start date
• The Matrix Utils Service loads and filters the prediction matrix for the requested cities.
• The Route Optimiser Service solves a weighted TSP using this matrix to compute the optimal visit order
• The Itinerary Handler Service schedules each leg according to the start date and requested stay durations
  • This scheduled route is also resolved into real, bookable flights or trains by this service
• A structured JSON itinerary is returned to the client

---

Stack

• Python
  • Google OR-Tools
  • NumPy
  • Requests
  • Uvicorn & FastAPI
• Amadeus Flights API
• Docker Compose

---

Services

API

• Publishes the /calculate_itinerary POST endpoint
• Takes an ItineraryRequest JSON model as a request body
• Synchronously calls subsequent services to generate different components of the optimised route
• Returns a JSONified bookable itinerary

ItineraryRequest Model

 class ItineraryRequest(BaseModel):
     # array of cities to be visited (not including the return to the city of origin at the end)
     selected_cities: List[str] = [] 
     # array of full days spent not travelling per city (including 0 at the end for return to city of origin)
     days_per_city: List[int] = []   
     # value assigned to each unit of time for cost/time matrix weighting
     time_weight: int
     # date on which travels should begin e.g. travel from depot city to city 1 on this date
     start_date: date 

Example /calculate_itinerary POST request

 curl -X POST http://localhost:8000/calculate_itinerary \
      -H "Content-Type: application/json" \
      -d '{
            "selected_cities": ["Berlin", "Prague"],
            "days_per_city": [0, 1, 0],
            "time_weight": 1,
            "start_date": "2026-05-11"
          }'

Example /calculate_itinerary response

{
  "bookable_legs": {
    "0": {
      "date": "2026-05-11",
      "dest": "PRG",
      "duration": 4.5,
      "mode": "train",
      "origin": "BER",
      "price": 30.0,
      "segments": [
        {
          "arrival": "2026-05-11",
          "departure": "2026-05-11",
          "from": "BER",
          "to": "PRG"
        }
      ]
    },
    "1": {
      "date": "2026-05-13",
      "dest": "BER",

      "duration": 4.5,
      "mode": "train",
      "origin": "PRG",
      "price": 30.0,
      "segments": [
        {
          "arrival": "2026-05-13",
          "departure": "2026-05-13",
          "from": "PRG",
          "to": "BER"
        }
      ]
    }
  },
  "metadata": {
    "end_date": "2026-05-13",
    "start_date": "2026-05-11",
    "total_cost": 60.0,
    "total_duration": 9.0
  }
}

Matrix Utils Service

• Saves updated flight data matrix / loads active flight data matrix into the Route Optimiser.
• Applies a filter to the flight data matrix to only load rows/columns relevant to selected cities, reducing data throughput for the Route Optimiser service.

Route Optimiser Service

• Computes the (predicted) optimal sequence of cities using Traveling Salesperson Problem (TSP) algorithms (using Google OR-Tools) with weighted cost-time matrices.
  • The flight data matrix serves as a 'prediction' that moving from city X to city Y is cheaper than city X to city Z, so for cities X, Y, Z, it predicts X→Y is cheaper than X→Z.
• Produces a preliminary route based on that prediction matrix including modes of transport (flight/train) and IATA codes.

Itinerary Handler Service

Itinerary Scheduler

• Converts preliminary optimised routes into date-constrained itineraries.
• Allocates days per city and applies scheduling rules to produce a realistic travel plan.
• Outputs a structured itinerary ready for booking.

Bookable Itinerary Service

• Maps scheduled itinerary legs to real-world flights and trains using available corridor data or fetching flights.
• Ensures each segment is bookable and respects travel constraints.
• Returns a fully actionable, structured travel plan in JSON format.

---

Current Features

• Multi-modal route optimisation across flights and trains.
• Weighted cost-time TSP solver for flexible prioritization.
• Date-constrained per-node scheduling to produce realistic itineraries.
• Mapping of abstract routes to bookable flights and train segments.

Future Features / Challenges in Development

• Potential Rail (ticketing) API access (thus far, none exist that are open to the public)
  • Many exist to tell you, "a train exists from A to B at Y time", but none that provide accurate ticketing/cost data. Many carriers also have extremely spotty coverage of this.
• Worldwide functionality
  • Theoretically, this will work right now - but computing a matrix for 19 destinations is already a large amount of our free Amadeus API quota and we can't really go much higher.
• New Flight API integrations
  • Amadeus shuts down self-service in July 2026. Enterprise still comes with a 2000 free allowance but incur a €0.025 charge per API call.
• It could be very worth storing historical flight matrices and training a regression model to predict future prices given variables (i.e. season, day of the week, etc.).

---